Online - OpenAccess - 978 3 7376 5093 9.OpenAccess

8th GACM Colloquium on
Computational Mechanics
For Young Scientists From Academia and Industry
August 28th – 30th , 2019
University of Kassel, Germany
Proceedings
Editors:
Tobias Gleim and Stephan Lange
The e-book is licensed under a Creative Commons Attribution-ShareAlike
4.0 International Public License.
Bibliographic information published by Deutsche Nationalbibliothek

The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data is available in the Internet at http://dnb.dnb.de.
ISBN 978-3-86219-5093-9
DOI: http://dx.medra.org/10.19211/KUP9783737650939
URN: https://nbn-resolving.org/urn:nbn:de:0002-450934
© 2019, kassel university press GmbH, Kassel

www.upress.uni-kassel.de
Preface
This conference book contains papers presented at the 8th GACM Colloquium on Computational
Mechanics for Young Scientists from Academia and Industry. The conference was held from
August 28th – 30th , 2019 in Kassel, hosted by the Institute of Mechanics and Dynamics of the department
for civil and environmental engineering and by the chair of Engineering Mechanics / Continuum Mechanics
of the department for mechanical engineering of the University of Kassel.
The aim of the conference is, to bring together young scientits who are engaged in academic and indus-
trial research on Computational Mechanics and Computer Methods in Applied Sciences. It provides a
plattfrom to present and discuss recent results from reseach efforts and industrial applications.
In more than 150 presentations, given by young scientists, current scientific developments and advances in
engineering practice in this field are presented and discussed. The contributions of the young researches
are supplemented by a poster session and plenary talks from four senior scientists from academia and
industry as well as from the GACM Best PhD Award winners 2017 and 2018.
We would like to thank the authors for their interesting contributions and the organizers of the 22
minisymposia for their effort making the 8th GACM Colloquium on Computational Mechanics
such a successful conference.
Kassel, August 2019
Tobias Gleim
Institute of Mechanics and Dynamics
Stephan Lange
Institute of Mechanics, Chair of Engineering Mechanics/Continuum Mechanics
iii
iv
Overview
GACM Best PhD Award Winner Lectures 1

1 The Mechanics of Brain Development
Silvia Budday
3 Kino–Geometric Modeling of Proteins and Molecular Mechanisms
Dominik Budday
7 Variational modeling of irreversible effects in cycled polycrystalline shape memory alloys
Johanna Waimann, Klaus Hackl and Philipp Junker
MS 01: Applications and Computational Methods in Metamaterials 11

11 Connection between micro- and macroscales in generalized mechanics
Bilen Emek Abali
15 On the modeling of microstructure influence
Arkadi Berezovski
19 A variational–based energy–momentum time integration of metamaterials in non-isothermal
rotordynamical systems
Michael Groß, Julian Dietzsch and Chris Röbiger
23 Elastic wave propagation in periodical one–dimensional problems: Accurate finite element
modeling
Radek Kolman, Sang Soon Cho, K.C. Park, Arkadi Berezovski, José González, Vítezslav
Adámek and Pert Hora
27 C0 continuous finite elements for gradient elasticity at finite strains
Johannes Riesselmann, Jonas Ketteler, Mira Schedensack and Daniel Balzani
31 Determining the Constitutive Parameters of a Macro–scale Second–Gradient Model for Planar
Pantographic Structures by Using Optimization Algorithms
Navid Shekarchizadeh and Masoud Abedi
35 Development of hybrid lattice materials using additive manufacturing
Christina Völlmecke, Vivian Dreßler, Daniel Melzer, Gregor Ganzosch and Anton Köllner
MS 02: Peridynamics and other non–Local Methods for Computational Mechanics 39

39 A finite strain microplane-plasticity model for concrete
Bobby Rio Indriyantho and Michael Kaliske
43 A nonlocal ductile damage model for high strength steels under dynamic multiaxial loading
conditions
Marvin Nahrmann and Anton Matzenmiller
47 Analysis of microwire torsion using a nonlocal dislocation based plasticity formulation
Kolja Zoller and Katrin Schulz
MS 03: Current Trends in Computational Fracture Mechanics 51

51 A phase-field model for mixed–mode fracture
Meng Fan, Thomas Wick and Yan Jin
55 Load Optimisation for Damage Reduction and Fracture Prevention
Fabian Guhr and Franz–Joseph Barthold
59 A phase–field fracture model based on Mindlin’s strain gradient theory
Resam Makvandi and Daniel Juhre
63 Modeling of interfacial crack propagation in strongly heterogeneous materials by using phase
field method
Thanh–Tung Nguyen, Julien Yvonnet, Daniele Waldmann and Qi–Chang He
67 Improved Durability Prediction of Cord–rubber Composites by the Material Force Method
Carmen Madalina Popa, Clemens Gebhardt and Michael Kaliske
71 Thermo–mechanical phase-field model for porous–ductile fracture
Jonathan Schulte, Maik Dittmann, Fadi Aldakheel and Christian Hesch
v
75 Thermo–Mechanical Coupling for Internal Thickness Extrapolation Elements
Felix Töller, Stefan Löhnert and Peter Wriggers
MS 04: Modern Trends in Numerical Continuation of Stationary Solutions and Application 79

79 On Solving Euler-Bernoulli Beams using a B-Spline based Representation for Random Vari-
ables
Christoph Eckert, Michael Beer and Pol D. Spanos
83 An Approach to Analyze the Stability of Quasiperiodic Motions with the Method of Charac-
teristics
Robert Fiedler and Hartmut Hetzler
87 Initialization of the continuation of stick-slip vibrations for a two dof friction oscillator
Jonas Kappauf and Hartmut Hetzler
MS 05: Multi-field problems in porous media mechanics 91

91 Thermo–hydro–mechanical modeling of liquid freezing in saturated porous media
Abdel Hassan Sweidan,Yousef Heider and Bernd Markert
97 Computational modelling of creep and rate dependent strength of freezing soils
Rodolfo Javier Williams Moises and Günther Meschke
MS 06: Modeling and simulation of fatigue in brittle and ductile materials 101
101 The FKM guidelines: Recommendations for fatigue life calculations of industrial components
Melanie Fiedler and Michael Vormwald
105 An Approach to Characterize Fatigue Crack growth by Material Forces
Jad Khodor and Michael Kaliske
109 Extension of the notch strain concept to transient cyclic material behaviour and residual
stresses
David Kühne, Martha Seiler, Franz Ellmer, Peter Hantschke and Markus Kästner
113 Transferability concepts of Fatigue Strength for high purity copper components in electrical
engines
Stefan Pruy, Markus Kästner and Peter Hantschke
117 Simulation of Fatigue with Cyclic Cohesive Zone Models
Stephan Roth, Meinhard Kuna and Björn Kiefer
121 Efficient phase–field modelling of fatigue in ductile materials
Martha Seiler, Thomas Linse, Peter Hantschke and Markus Kästner
125 High–order cycle jump integration of a concrete fatigue damage model
Thomas Titscher, Jörg F. Unger and Javier Oliver
MS 07: Higher Order Numerical Methods in Space and Time 129

129 Mixed finite element formulations and energy–momentum time integrators for thermo–mechanically
coupled fiber–reinforced continua
Julian Dietzsch, Michael Groß and Lars Flessing
133 Application of diagonally implicit Runge Kutta methods to structural dynamics problems
Matthias Grafenhorst and Stefan Hartmann
139 A higher–order energy–momentum scheme for a non–isothermal two–phase dissipation model
of fibrous composites
Michael Groß, Julian Dietzsch and Chris Röbiger
143 Implicit Runge–Kutta Schemes and Galerkin Me.thods Similarities and Differences
Bettina Schröder and Detlef Kuhl
147 Higher order reciprocal mass matrices via heuristic customization
Anton Tkachuk
MS 08: Smart, Active and Polymer materials 151
vi
151 Modeling the thermomechanical behavior of strain–induced crystallization in unfilled poly-
mers
Serhat Aygün and Sandra Klinge
155 Crystallization of semi–crystalline polymers simulated with a cellular automaton
Tobias Daniel Horn, Hans Wulf and Jörn Ihlemann
159 Towards thermo–viscoelastic experimental characterisation and numerical modelling of VHB
polymer
Mokarram Hossain and Zisheng Liao
163 A Macroscopic Model for Magnetorheological Elastomers based on Microscopic Simulations
Karl A. Kalina, Philipp Metsch, Jörg Brummund and Markus Kästner
167 Experimental Characterization, Modeling and FE-Simulation of Curing Phenomena in an
Elastomeric Adhesive
Ralf Landgraf and Jörn Ihlemann
171 Modeling the Chemo-Thermo-Mechanical Material Behavior of an Epoxy Resin System
Chris Leistner and Stefan Hartmann
175 Modelling and simulation approaches for the numerical analysis of UV curing polymers used
in additive manufacturing
Thomas Rehbein, Alexander Lion and Michael Johlitz
179 Simulation of delamination in ferromagnetic–ferroelectric laminate composites
Alexander Schlosser and Andreas Ricoeur
183 Simulation and Modeling of Stator Elastomers in Positive Displacement Motors
Alexander Schöne, Ralf Landgraf and Jörn Ihlemann
187 Transient dissipation heating in polycrystalline ferroelectrics accounting for mutual nonlinear
couplings
Andreas Warkentin and Andreas Ricoeur
MS 09: Advances in full–field strain and temperature measurements for application in experi-
mental mechanics 191
191 Numerical treatment of material parameter identification using finite elements
Rose Rogin Gilbert and Stefan Hartmann
195 Parameter identification for constitutive models of textile composite materials using digital
image correlation
Justin Felix Hofmann, Claudia von Boyneburgk, Sophie Tunger, Hans–Peter Heim and Detlef
Kuhl
199 Monitoring of multiaxial failure surfaces and the evolution of deformation bands in PVC
foams using 3D digital image correlation
Anne Jung, Francesca Concas and Stefan Diebels
203 Full–field strain measurements on single struts of a Ni/PU hybrid foam
Martin Reis, Stefan Diebels and Anne Jung
MS 10: Dynamic Systems with Non–standard Constraints: Simulation, Control and Numerical
Analysis 207
207 A Lie group generalization of the RATTLE scheme applied to nonholonomic constraints
Stefan Hante
211 A Brief Survey on Non-standard Constraints: Simulation and Optimal Control
Dominik Kern and Ulrich J. Römer
215 Comparison of timestepping schemes for non–smooth mechanics using orbital convergence
Manuela Paschkowski and Martin Arnold
MS 12: Computational Multiscale and Multiphysics Simulations of Microheterogeneous Materials219

219 Investigation of the fractal properties of aerogels by diffusion–limited aggregation models
Rasul Abdusalamov, Mikhail Itskov, Barbara Milow, Gudrun Reichenauer and Ameya Rege
vii
223 Simulation of Dual–Phase Steel Microstructures Using Optimized Arrangements of Finite
Cells
Yannick Francis Fangye, Niklas Miska and Daniel Balzani
227 Modeling interpenetrating metal matrix composites using a fast Fourier transform formulation
Dominik Horny and Katrin Schulz
231 Can the constituent stress–strain laws of composite materials be determined uniquely from
integral measurements?
Fabian Sewerin
235 Numerical thermo–elasto–plastic analysis of residual stresses on different scales during cooling
of hot formed parts
Sonja Uebing, Lisa Scheunemann, Dominik Brands and Jörg Schröder
MS 13: Advanced Numerical Methods for Fluid-Structure Interaction 239

239 A weak Dirichlet–Neumann Partitioned Algorithm for 1–Dimensional Structures Embedded
in a 3–Dimensional Flow
Nora Hagmeyer, Matthias Mayr and Alexander Popp
243 Spline–Based Finite Element Method for Exact Geometric Interface Representation
Michel K.P. Make, Norbert Hosters and Marek Behr
MS 14: Computational Contact & Interface Mechanics 247

247 An interface element formulation for frictionless normal contact between rough surfaces
Jacopo Bonari, Maria Rosaria Marulli, Nora Hagmeyer, Matthias Mayr, Alexander Popp and
Marco Paggi
251 Multiscale Rubber Friction Homogenization on Rough Surfaces
Felix Hartung and Michael Kaliske
255 An unbiased self–contact formulation for explicit FEA stabilized by the bipenalty method
Ján Kopačka, Dusan Gabriel and Radek Kolman
259 Extended Mortar Methods for Isogeometric Analysis
Melanie Krüger and Christian Hesch
263 Solving computational contact problems efficiently at large scale
Matthias Mayr and Alexander Popp
267 Isogeometric Dual Mortar Methods for Computational Contact and Interface Mechanics
Alexander Popp, Linus Wunderlich, Alexander Seitz and Barbara Wohlmuth
271 A residual–type a posteriori estimator for the adaptive discretization in space and time of a
viscoelastic contact problem
Mirjam Walloth
MS 15: Novel Formulations and Discretization Methods for Thin–walled Structures 275
275 Molecular Interactions of Slender Fibers Modeled as Cosserat Continua
Maximilian J. Grill, Wolfgang A. Wall and Christoph Meier
279 An isogeometric scaled boundary shell formulation for thin–walled structures
Markus Klassen and Sven Klinkel
283 Free Material Optimization of Multilayer Composite Materials
Simon Loske and Franz–Joseph Barthold
287 Mesh Tying and Contact Algorithms for Nonlinear Beam-to-Solid Interactions
Ivo Steinbrecher and Alexander Popp
291 Analysis of support conditions applying coupling-based deformation modes in Generalised
Beam Theory (GBT)
Fabiola Tartaglione, Marcelo J. Bianco, Abinet K. Habtemariam, Volkmar Zabel and Carsten
Könke
MS 16: Recent Advances on Time Integration with Focus on CFD Applications 295
viii
295 Multirate GARK schemes for the simulation of fluid–structure interaction: Linear stability
and the geometric conservation law
Sascha Bremicker-Trübelhorn and Sigrun Ortleb
299 Stability Analysis of IMEX–DG–Schemes for Advection–Diffusion with Focus on the Diffusion
Treatment
Sigrun Ortleb
303 Comparison of exponential–explicit (EXPEX) schemes in the domain–based implicit–explicit
(IMEX) setting with IMEX–RK schemes for CFD applications
Veronika Straub and Sigrun Ortleb
307 High order IMEX splitting methods for low Mach number flows
Jonas Zeifang
MS 17: Coupled Problems 311

311 Computation of the Stability Limit of Nonlinear Bladerow Flutter with a Fully Coupled
Frequency Domain FSI Solver
Christian Berthold, Johann Gross, Christian Frey and Malte Krack
315 Adaptive Multilevel Optimization of Fluid–Structure Interaction Problems
Johanna Katharina Biehl and Stefan Ulbrich
319 Finite element modelling of ultrasonic guided waves propagation in the fluid–composite struc-
ture interaction
Bhuiyan Shameem Mahmood Ebna Hai and Markus Bause
323 An efficient implementation of gradient-enhanced crystal plasticity models
Volke Fohrmeister and Jörn Mosler
327 A thermomechanical finite element based framework for phase transitions in laser based ad-
ditive manufacturing processes
Isabelle Guschke, Thorsten Bartel and Andreas Menzel
331 A Sliding Interface Approach with Application to Twin–Screw Extruders
Jan Helmig and Stefanie Elgeti
335 Partitioned coupling of fluid–structure interaction for the simulation of floating wind turbines
Jorrid Lund, Bjarne Wiegard, Lars Radtke, Marcel König and Alexander Düster
339 Particle Finite Element Simulation for Additive Manufacturing Processes of Fresh Concrete
Janis Reinold, Jithender Jaswant Timothy and Günther Meschke
343 Optimisation of Diffusion Driven Degradation Processes
Navina Waschinsky, Franz–Joseph Barthold and Andreas Menzel
347 Acoustic–structure interaction in the Scaled Boundary Finite Element Method for prismatic
geometries
Paul Wasmer, Jannis Bulling, Hauke Gravenkamp and Jens Prager
351 Modelling the Soil–Water Interaction at Fluid–Infiltrated Roads under Loading of Steady
State Rolling Tires
Ines Wollny, Wai Ching Sun and Michael Kaliske
MS 18: Crack initiation and crack growth in modern composite materials: fracture of interfaces,
material discontinuities and composite laminates 355
355 Constitutive modelling of carbon fibre reinforced polymers for adhesively bonded cfrp–steel
joints
Michael Donhauser and Anton Matzenmiller
359 Hyperelastic adhesive joints: from crack nucleation to ultimate failure
Julian Felger and Wilfried Becker
363 Hyperelastic adhesive joints: from crack nucleation to ultimate failure
Philipp L. Rosendahl, Yves Staudt, Alexandra P. Schneider, Jens Schneider and Wilfried
Becker
367 J–integral and mixed–mode cohesive zones
Johannes Scheel and Andreas Ricoeur
ix
371 Finite element model to describe crack deflection and constrained micro–cracking
Markus Tauscher, Martin Pletz, Florian Arbeiter and Clara Schuecker
375 Stress Control on Interfaces Using Shape Optimisation
Felix Wohlgemuth and Franz–Joseph Barthold
MS 19: Brittle Fracture Analysis in Solids: Experiment and Simulation 377

377 Fragment Detection Algorithm for Postprocessing High Velocity Impact and Natural Frag-
mentation
Marvin Becker, Marina Seidl, Miriam Mehl and Mhamed Souli
381 Multiphase-field Simulation Study of Brittle Anisotropic Crack Propagation in Sandstones
Michael Späth, Nishant Prajapati, Christoph Herrmann, Daniel Schneider, Michael Selzer
and Britta Nestler
MS 20: Numerical methods in biomechanics 385

385 User–defined material subroutines for FEM calculations based on automatic differentiation
of strain energy functions
Markus Hillgärtner, Theron Guo and Mikhail Itskov
389 Nonlinear fibre distribution in a finite element model of a human left ventricle for sparse and
non–constant data sets
David Holz, Minh Tuan Duong, Kim Myoungsic, Muhannad Alkassar, Sven Dittrich and
Sigrid Leyendecker
393 A Frictional and Adhesive Contact Model for Debonding of the Bone–Implant interface Based
on State Variable Friction Laws
Katharina Immel, Thang X. Duong, Vu–Hieu Nguyen, Guillaume Haïat and Roger A. Sauer
397 Finite element technologies for stent simulations
Kiran Manjunatha, Jan Frischkorn, Stefanie Reese
401 Numerical simulation of the viral entry into a cell by receptor driven endocytosis
Tillmann Wiegold, Sandra Klinge, Robert P. Gilbert and Gerhard A. Holzapfel
405 Optimization–based comparison of different anisotropic models for growth and remodeling in
arterial walls
Anna Zahn and Daniel Balzani
MS 21: Representative Volume Elements: Microstructure Analysis and Homogenization 409

409 Application of Voronoi tessellations to model cellular aerogels
Rajesh Chandrasekaran, Markus Hillgärtner, Ameya Rege, Barbara Milow and Mikhail Itskov
413 Microstructural convergence analysis as basis for an efficient and accurate two–scale FE–
FFT–based simulation approach
Christian Gierden, Julian Kochmann, Johanna Waimann, Bob Svendsen and Stefanie Reese
417 Large strain two–scale simulations on laptop computers
Oliver Kunc and Felix Fritzen
421 Thermal Homogenization Using Image Data
Julian Richard Michael Lißner and Felix Fritzen
425 Numerical and Experimental Characterization of Additively Manufactured Mild Steel under
Tension
Jorge Lizarazu, Luise Göbel, Stefan Linne, Susanne Kleemann and Tom Lahmer
429 A Homogenization Method with Minimal RVE Boundary Conditions
Maximilian Müller and Friedrich Gruttmann
433 Cauchy–Born rule based continuum simulation of the mechanical properties of single–layered
graphene sheets
Qige Zheng and Jens Wackerfuß
437 Shear Deformations in a Multiscale Shell Formulation
Jan Zoller and Friedrich Gruttmann
x
MS 22: Recent Trends in Phase–Field Modeling: Theory, Numerics, Applications 441
441 An efficient matrix–free finite element solver for the Cahn–Hilliard equation
Fabian Castelli and Willy Dörfler
445 Improving the numerical solution of the phase–field equation by the systematic integration of
analytic properties of the phase–field profile function
Michael Fleck, Felix Schleifer, Markus Holzinger and Uwe Glatzel
451 Phase–field damage model coupled with ageing in polymer adhesives due to moisture diffusion
Siva Pavan Josyula and Stefan Diebels
455 An anisotropic phase–field model based on a Representative Crack Element (RCE)
Dennie Supriatna, Johannes Storm and Michael Kaliske
459 Reconstruction of a Ferroelectric Stripe Domain Array due to Charge Defect Migration
Ivan Vorotiahin, Anna Morozovska and Yuri Genenko
463 Phase–field simulation of the flexoelectricity in ferroelectrics
Xian–Dong Zhou and Bai–Xiang Xu
MS 23: Non–Newtonian Fluids 465

465 Comparison of different rheological models in pressure–hole error simulation
Hogenrich Damanik and StefanTurek
469 Shape Optimization in Polymer Extrusion Using Shear–Thinning Fluids
Sebastian Eusterholz and Stefanie Elgeti
473 Heat Dependent Thermo–Viscous Shear–Thinning Fluid Flowing around a Double–Sphere
Configuration within a Pipe
Markus Rütten
Authors Index 479
xi
xii
GACM Best PhD Award Winner Lectures
Proceedings of the 8th GACM Colloquium on Computational Mechanics

for Young Scientists from Academia and Industry
August 28–30, 2019 in Kassel, Germany
The Mechanics of Brain Development

∗ 1
Silvia Budday
1 Friedrich-Alexander-University Erlangen-Nürnberg, Erlangen, Germany
Abstract. The characteristically folded surface morphology is a classical hallmark of the mammalian brain.
During development, the initially smooth surface evolves into an elaborately convoluted pattern, which
closely correlates with brain function and serves as a clinical indicator for pathological conditions. Here,
we combine computational and experimental analyses to show that physical forces control pattern selec-
tion in the developing brain. We closely consider cellular processes during brain development to estab-
lish a mechanical model for brain growth. The model consists of a morphogenetically growing outer cor-
tex and a stretch-induced growing inner core. Our results demonstrate that mechanical instabilities can
explain cortical folding during brain development. Combining physics and biology holds promise to ad-
vance our understanding of human brain development, to enable early diagnostics of cortical malforma-
tions, and to improve treatment of neurodevelopmental disorders such as epilepsy, autism, and schizophrenia.
Introduction
Neurodevelopment involves a highly orchestrated sequence of events, which is tightly regulated by the
complex interplay of various cell types and molecular factors both in space and time [2]. The first
phase of brain development is characterized by cell division and cell migration, ultimately resulting
in the formation of the outer cell-rich brain layer, the cortex. At the end of the neuronal migration
process, its surface is still smooth. During the second phase of brain development, cortical neurons begin
to interconnect and form synapses. This connectivity process results in an excessive expansion of the
cortex, which is constrained by the much less growing inner core. Therefore, compressive stresses build
up in the cortex, which eventually induce cortical folding as a mechanism to release the growth-induced
compressive stresses [1]. Physics-based modeling can help correlate neurodevelopmental disruptions on
the cellular level with structural malformations on the organ level [2]. The latter can easily be detected
through modern imaging techniques.
Computational model for brain growth
We model brain growth using the nonlinear field theories of mechanics supplemented by the theory of
finite growth [1]. We multiplicatively decompose the deformation gradient into an elastic and a growth
contribution F = F e · F g . We approximate the behavior of brain tissue using a hyperelastic neo-Hookean
material model with the strain energy function ψ = 1/2 µ[ F e : F e − 3 − 2 ln J e ] + 1/2 λ ln2 J e , expressed
in terms of the elastic components of the deformation gradient F e and the Jacobian J e = detF e . The
Piola stress then yields P = P e ·F g−T , where P e = ∂ψ/∂F e . In the absence of volume forces, the balance
.
of linear momentum reduces to DivP = 0, which we solve numerically using the finite element method.
We close the set of nonlinear equations by prescribing the growth contribution F g and its evolution
in time based on the cellular processes during brain development [1]. We assume that brain tissue
expands isotropically and express F g in terms of a single scalar valued growth multiplier ϑ, F g = ϑ1.
To mimic the connectivity process during the second phase of brain development, we let the cortex grow
morphogenetically in time at the constant growth rate Gc . We further assume that subcortical growth
is stretch-induced (see Figure 1a), as axons grow when mechanical tension is applied. Growth is only
activated if the elastic volume stretch J e exceeds its baseline value J 0 .
∗ Corresponding author: silvia.budday@fau.de ORCID iD: 0000-0002-7072-8174
1
Figure 1. a) Mechanical model for brain growth. Simulation of a simplified brain geometry (b) and swelling
experiments on elastomers (c) show the emergence of complex wrinkling patterns similar to those of
3D-printed brain models extracted from magnetic resonance images at different stages of human brain
development, gestational week 22 and term (d), as well as at the age of 30 (e).
Experimental model for brain growth
We experimentally mimic brain folding by using soft elastomeric gels, which swell superficially when
immersed in solvents [3]. We cast a simplified initial geometry of the inner core using a soft elastomer
(Sylgard 184, Dowsil, mixed in a 45:1 mass ratio) and cure it at 75◦ C for about 30 min. After removal
from the mold, the cortex layer (Sylgard 184, mixed in a 35:1 mass ratio) with blue color pigment is
drop-casted, wherein each layer is cured at 75◦ C for about 20 min until the desired layer thickness is
reached. The whole specimen is finally cured over night. To induce growth (or swelling) of the outer
cortical layer, the elastomer-brain is immersed in solvent (Hexanes, Sigma-Aldrich) at room temperature
for about two hours, as illustrated in Figure 1c.
Results
Figures 1d and 1e show 3D-printed human brain models extracted from magnetic resonance images at
different stages of development: at gestational week 22, when the cortex is still smooth (d); at term, when
all major fissures have formed (d); at the age of 30, when the brain has grown to its maximum size (e).
Figures 1b and 1c impressively demonstrate that both the fairly simple mechanical model and the swelling
elastomers result in complex wrinkling patterns that well resemble the morphology of mammalian brains.
Conclusion
The current work strongly suggests that mechanical instabilities are the underlying mechanism of cortical
folding during brain development, and that mechanical factors play an important role in pattern selec-
tion. The mechanical model bridges the scales from cellular events such as neuronal connectivity and
axon elongation to clinically relevant characteristics on the macroscopic scale, e.g. abnormal gyrification
patterns, which are accessible through modern imaging techniques such as magnetic resonance imaging.
Acknowledgments
The work by Alexander Greiner and the financial support by the German Research Foundation grant
STE 544/50 and the Emerging Talents Initiative (ETI) by the FAU are gratefully acknowledged.
References
[1] Silvia Budday, Paul Steinmann, and Ellen Kuhl. The role of mechanics during brain development.
J. Mech. Phys. Solids 72 (2014), pp. 75–92.
[2] Silvia Budday, Paul Steinmann, and Ellen Kuhl. Physical biology of human brain development.
Front. Cell. Neurosci. 9 (2015), p. 257.
[3] Tuomas Tallinen, Jun Young Chung, François Rousseau, Nadine Girard, Julien Lefèvre, and L Ma-
hadevan. On the growth and form of cortical convolutions. Nat. Phys. 12.6 (2016), pp. 588–593.
2

Kino-Geometric Modeling of Proteins and Molecular

Mechanisms
∗1
Dominik Budday
1 formerly at the Chair of Applied Dynamics, Friedrich-Alexander-Universität Erlangen-Nürnberg, 91058 Erlangen, Germany
Abstract.
Proteins are dynamic biomolecules that perform an enormous variety of cellular functions on a broad range
of spatio-temporal scales. Their conformational ensemble is a crucial determinant of functionality in health
and disease. While experimental and computational advances have increasingly enabled atomically detailed
insights into protein molecular mechanisms, the need for efficient, yet elaborate integrative computational
methods to resolve functional motions across scales remains considerable. Here, we lay out a robotics-inspired,
kino-geometric model that efficiently captures small- and large-scale collective motions, with dihedral angles as
torsional degrees of freedom and non-covalent interactions as constraints. Using geometric tools, we decompose
the biomolecules into rigid substructures and coordinated motions between them. We develop an efficient,
high-dimensional motion planning algorithm dCC-RRT that integrates the principle of minimal frustration via
dynamic, Clash-avoiding Constraints (dCC) into a rapidly-exploring random tree (RRT). Shown exemplarily
on the enzyme Adenylate Kinase, the algorithm closely approximates its conformational transition and agrees
well with motions from Molecular Dynamics simulations and experimental data, thereby providing a structural
basis for allosteric networks that drive conformational change. Overall, our kino-geometric modeling approach
is a robust and efficient alternative to obtain high-level insights into molecular mechanisms across scales, with
broad applications in protein engineering, drug design, and human health.
Introduction
Proteins perform a tremendous variety of biological functions. Synthesized by the ribosome in our cells,
they fold into rather complex, three-dimensional shapes termed their native conformation. This state,
together with a subtle interplay between stability and flexibility, allows them to interact with various
binding partners, such as small ligands, nucleotides or other proteins. The free-energy surface that char-
acterizes different protein states and underlies their structural rearrangements is a critical determinant for
protein functionality. Thus, efficiently and reliably characterizing the temporal and spatial distributions
of protein conformational ensembles is crucial to promote our understanding of molecular mechanisms in
health and disease. Although typical Molecular Dynamics (MD) based methods can provide atomically
detailed trajectories, they require tremendous computational resources and tend to face limitations re-
garding the vast scales in space and time that are covered by protein motion. As shown in Fig. 1, these
range from pico-second bond length and angle fluctuations, over side-chain flips and loop motion to rear-
rangements of individual protein domains or entire chains that may occur on the order of milliseconds or
seconds. To mitigate these limitations, we take an alternative, time-independent approach via kinematics
based analyses and conformational sampling towards high-level insights into the molecular mechanisms
of protein function.
KGS – our Kino-Geometric Sampling framework
Originating in traditional robotics, our Kino-Geometric Sampling (KGS) and modeling framework effi-
ciently captures small- and large-scale collective motions important for molecular function [2, 3, 4, 5]. As
shown in Fig. 2, we model the dihedral angles φ, ψ, and χ of main- and side-chain covalent single bonds
as d torsional degrees of freedom q ∈ Td of the molecule. Those are coupled by hydrogen bonds, disulfide
∗ Corresponding author: dominik.budday@fau.de ORCID iD: ORCID 0000-0002-6180-2125
3
Figure 1. Spatio-temporal scales of protein motions and appropriate computational methods for their analysis.
Figure 2. Protein fragment represented as a kinematic spanning tree, with dihedral angles (φ, ψ, χ) as degrees
of freedom between rigid body vertices (individually colored) and non-covalent interactions such as
hydrogen bonds (yellow dashes) encoded as geometric constraints that form closed kinematic cycles
with a left (L) and right (R) branch leaving from a common anchor.
bonds, or hydrophobic forces to move in a coordinated fashion, while bond lengths and angles are fixed.
We integrate these non-covalent interactions (yellow dashes) as cycle closure constraints Φ(q), leading to a
description of the molecular conformation space in terms of a constraint variety Q = {q ∈ Td | Φ(q) = 0}.
Introducing the constraint Jacobian J, we obtain instantaneous (velocity) constraints
(
dΦ = 0, if q̇ constraint observing
= Jq̇
dt 6= 0, if q̇ constraint violating
which characterize two disjoint, orthogonal subspaces. In this fashion, we can focus on the collective
motions of dihedral angles that lie in the nullspace of J and preserve the global constraint network
[3] (i.e. Jq̇ = 0), or the ones that lie in the range of the transposed Jacobian range(JT ) and perturb
constraints in a hierarchical manner [2, 5]. Coupled to sophisticated motion planning strategies, this
basis allows for efficient sampling of conformational pathways between different molecular states.
4
Figure 3. Exciting insights into the molecular mechanisms of ADK using Kino-Geometric Modeling and Sampling.
Results and Discussion
KGS provides insights into molecular flexibility and stability on various levels. To graphically display these
capabilities, we make use of the well-studied example protein Adenylate Kinase (ADK), an important
enzyme that is associated with maintaining energy levels in our cell and performs rather large opening
and closing motions throughout its catalytic cycle (see e.g. [1]). Based on a deposited molecular structure
from the protein data bank (PDB) (Fig. 3, top center, PDB code 4ake), KGS automatically generates a
kinematic spanning tree of the molecule and identifies significant non-covalent interactions to include as
constraints, based on typical geometric and energetic criteria [6] used in the protein rigidity community
(see [9] for a broader review on rigidity theory applied to macromolecules). Performing geometric rigidity
analysis [3] on this graph, we identify larger rigid substructures that emerge when only motions that
completely respect constraints are accessible, i.e., the motions in the nullspace of J. The top left panel in
Fig. 3 displays those substructures in ADK individually colored, demonstrating relatively rigid secondary
structure helices that stabilize the enzyme domains, but still allow for its characteristic opening and
closing motions [1, 2]. By admitting motion modes that increasingly perturb constraints in a hierarchical
fashion, we observe global motions that span the entire protein (Fig. 3, top right) [5]. The overall mobility
pattern, colored by increasing atomic root-mean squared fluctuations (RMSF) from blue to red matches
the enzyme domains fairly well: the stable CORE domain (mostly blue), and the more flexible LID (left)
and NMP (right) binding domains (green/red) [1]. Both types of analyses are instantaneous in nature
and capable to report these stability and flexibility measures in a matter of seconds.
In order to gain insights into the molecular mechanisms of activation and identify pathways between
different states of the same molecule, we have integrated various sophisticated motion planning strategies
in KGS. These algorithms are based on rapidly exploring random trees (RRT) [10] or Poisson planning
[7], and specifically integrate the principal of minimal frustration to automatically explore energetically
accessible regions of conformation space [4, 8]. This is achieved by forming dynamic Clash-avoiding
Constraints (dCC) whenever two atoms are in close contact, biasing exploration towards more favorable
directions. Speaking in terms of the constraint variety Q, this approach dynamically alters its dimensions
by adding constraints, allowing us to perform manifold hopping during exploration and exploiting motion
where accessible. For our transition toolbox with dCC-RRT, we focus on constraint observing motions
restricted to the (dynamically changing) nullspace of J and only include constraints from hydrogen
bonds present in both conformations of interest. Providing the closed form of ADK as a second input
(Fig. 3, bottom center, PDB code 1ake), KGS identifies a connective transition pathway between the
open and the closed state (Fig. 3, bottom left) in a few hours [4]. Comparing RMSF from the dCC-
RRT pathway to an MD trajectory (Fig. 3, center left), we observe striking qualitative and quantitative
5
agreement [2]. A further comparison with unbiased, unguided sampling around its open state (blue
line) demonstrates very similar trends, but overall much less amplitude in motion. Clearly, cleverly
designed motion planning is necessary to navigate around obstacles and widely cover conformation space.
Together with our underlying KGS model, we capture similar molecular characteristics as MD, but at
a much lower computational cost. On top of that, we have shown that our exploration pathways pass
several experimentally determined intermediate states of ADK, further validating our motion planner
dCC-RRT [4]. Clash constraints not only guide exploration towards sterically favorable conformations,
but at the same time inform on residue contacts along the sampled pathways. By monitoring the list of
pairwise dCC, we identify residue interaction networks that can propagate collective motion over large
distances in the molecule. For ADK, the largest network shown in red spans all three protein domains
and seems to be a key component for enabling its conformational change (Fig. 3, bottom right).
Overall, our results establish KGS as an efficient and insightful alternative to study molecular mechanisms
across scales. As part of a larger computational pipeline for protein and nucleotide dynamics, the various
results from our toolbox may serve as input to fine-tune analysis via more detailed, but costly methods
such as MD, with wide applications in protein design and drug development. Source code, analysis scripts
and several examples are freely available for download at https://github.com/ExcitedStates/KGS.
Acknowledgments
I would like to express my deep thanks to Prof. Dr.-Ing. habil. Sigrid Leyendecker and Associate Adjunct
Prof. Dr. Henry van den Bedem for their continuous support and valuable guidance during my exciting
time as doctoral candidate, both at the FAU in Germany as well as the SLAC National Accelerator
Laboratory in the US. I gratefully acknowledge financial support by BaCaTeC (project number 7, 2014 −
1), the DFG (project number 401512690), and the Deutsche Telekom Stiftung.
References
[1] Oliver Beckstein, Elizabeth J Denning, Juan R Perilla, and Thomas B Woolf. Zipping and unzipping
of adenylate kinase: atomistic insights into the ensemble of open ↔ closed transitions. J. Mol. Biol.
394.1 (2009), pp. 160–176.
[2] Dominik Budday. “High-Dimensional Robotics at the Nanoscale - Kino-Geometric Modeling of Pro-
teins and Molecular Mechanisms”. PhD thesis. Friedrich-Alexander-Universität Erlangen-Nürnberg,
2019.
[3] Dominik Budday, Sigrid Leyendecker, and Henry van den Bedem. Geometric analysis characterizes
molecular rigidity in generic and non-generic protein configurations. J. Mech. Phys. Solids 83 (2015),
pp. 36–47.
[4] Dominik Budday, Rasmus Fonseca, Sigrid Leyendecker, and Henry van den Bedem. Frustration-gui-
ded motion planning reveals conformational transitions in proteins. Proteins: Struct. Funct. Bioinf.
85.10 (2017), pp. 1795–1807.
[5] Dominik Budday, Sigrid Leyendecker, and Henry van den Bedem. Kinematic Flexibility Analysis:
Hydrogen Bonding Patterns Impart a Spatial Hierarchy of Protein Motion. J. Chem. Inf. Model.
58.10 (2018), pp. 2108–2122.
[6] Bassil I Dahiyat, D Benjamin Gordon, and Stephen L Mayo. Automated design of the surface
positions of protein helices. Protein Sci. 6.6 (1997), pp. 1333–1337.
[7] Rasmus Fonseca, Dominik Budday, and Henry van den Bedem. Collision-free poisson motion plan-
ning in ultra high-dimensional molecular conformation spaces. J. Comput. Chem. 39.12 (2018),
pp. 711–720.
[8] Amélie Héliou, Dominik Budday, Rasmus Fonseca, and Henry van den Bedem. Fast, clash-free RNA
conformational morphing using molecular junctions. Bioinformatics 33.14 (2017), pp. 2114–2122.
[9] Susanne MA Hermans, Christopher Pfleger, Christina Nutschel, Christian A Hanke, and Holger
Gohlke. Rigidity theory for biomolecules: concepts, software, and applications. WIREs Comput.
Mol. Sci. 7.4 (2017), e1311.
[10] Steven M LaValle. Rapidly-exploring random trees: A new tool for path planning. Tech. rep. 98-11.
Iowa State Univ., 1998.
6

August 2830, 2019 in Kassel, Germany
Variational modeling of irreversible eects in cycled

polycrystalline shape memory alloys
Johanna Waimann
∗ 1, Klaus Hackl
2 and Philipp Junker
3
1 Institute of Applied Mechanics, RWTH Aachen University, Germany

2 Institute of Mechanics of Materials, Ruhr-Universität Bochum, Germany
3 Chair of Solid Mechanics, University of Wuppertal, Germany
Abstract. The material behavior of shape memory alloys (SMAs) is characterized by a diusionless solid/solid
phase transformation between an austenitic and several martensitic phases. Compared to transformations
in classical steel, the transformation process is not only induced by thermal loading but also by mechanical
loads. Due to the related properties and namely the eects of pseudoelasticity and pseudoplasticity, SMAs are
very attractive materials for industrial applications. The presented work focuses the functional fatigue of this
special class of materials which occurs during cyclic loading of the specimen. Three models are presented which
are based on the principle of the minimum of the dissipation potential (PMDP) and take into account dierent
aspects of the complex material behavior: a simultaneous plastic deformation, the formation of irreversible
martensite during mechanical cycling and the fatigue during thermal cycling. In addition, several simulations
show the functionality of the material models.
Introduction
Cyclic tension tests of SMAs show that due to some formed dislocations (cf. [5]) the material is func-
tionally fatigued: a remaining strain is accumulating, the transformation from austenite to martensite is
favored and the reverse transformation is delayed (cf. [3]). During thermal cycling the opposite eect
takes place: the transformation from austenite to martensite is delayed and the back transformation is
favored (see [1]).
The material models in the following sections are also presented in [6] and show dierent aspects of the
microstructural evolution. Whereas in the rst presented model the phase transformation is coupled
to a formation of irreversible martensitic volume fractions, which are the direct results of the formed
dislocations (see [5]), the second model displays the fatigue phenomena by considering a simultaneous
plastic deformation. Thereby, a coupled dissipation function is introduced to take into account the in
parallel evolving microstructural processes. The third model is additionally able to display the material's
fatigue performance during thermal cycling by storage of the material's transformation history. The
polycrystalline character of the examined shape memory alloys is additionally considered by an evolving
orientation distribution function [4]. Each model's description is accompanied by numerical results which
proof its performance.
1 Simultaneous formation of irreversible martensite
Based on the principle of the minimum of the dissipation potential, the rst material model is taking into
account a formation of irreversible martensite during the phase transformation between one austenitic
and n martensitic phases. The transformation processes are considered by two internal variables: the
reversible and the irreversible volume fractions, λ and ρ. To additionally take into account the poly-
crystalline character of the examined SMAs, a third internal variable is used. The set of Euler angles
α = {ϕ, ϑ, ω} describes the averaged orientation of the transforming grains based on [4]. Compared to a
∗ Corresponding author: waimann@ifam.rwth-aachen.de ORCID iD: 0000-0002-7579-7240
7
grain-wise determination of the volume fractions, it enables a much faster calculation of the microstruc-
tural evolution. The used variational method, the PMDP (cf. [2]), is based on the idea of formulating a
Lagrange function

L = Ψ̇ (ε, θ, α, λ, ρ) + D λ, ρ, α̇, λ̇, ρ̇ + cons → min , (1)
α̇,ρ̇,λ̇
which includes the rate of the Helmholtz free energy Ψ, a dissipation function D and constraints which
have to be fullled and are considered by Lagrange and Kuhn-Tucker parameters. The quantities ε and
θ are the total strain and the temperature. The following minimization of L with respect to the rates of
the internal variables directly results in evolution equations for them. Considering the microstructural
eects, the Helmholtz free energy has the form
1
Ψ= (ε − η̄ R − ῡ R ) : Ē : (ε − η̄ R − ῡ R ) + c̄ (2)
2
with the eective quantities
n
X n
X
η̄ R = QT · λi η i · Q ῡ R = QT · ρi η i · Q
i=0 i=0
" n #−1 n
X −1
X
Ē = (λi + ρi ) (Ei ) c̄ = (λi + ρi ) ci . (3)
i=0 i=0
In Equations (2) and (3), η i is the transformation strain of phase i, which is an experimentally observed
quantity. Related to the irreversible volume fraction, a part of the transformation strain remains in the
material, which is given by υ i = ρi η i . By use of the rotation tensor Q(α), the transformation strains
are rotated in the energetically optimized averaged orientation of the transformations. The quantities Ei
and ci describe the stiness and the caloric energy of the individual phases i. The in Equation (1) used
function D denotes the energy which dissipates due to the change of the microstructure. With use of the
dissipation parameter rT1 , the coupling parameters f and g as well as the viscous parameter rR , it has
the form v
u n 2 X √
2 rR 2
n
uX
D = rT1 t
2
f λ̇i + (g ρ̇i ) + ϕ̇ + ϑ̇2 + 2ϕ̇ω̇ cos ϑ + ω̇ 2 (4)
i=0 i=0
2
and includes a rate-independent coupled approach for the part related to the transformations and a
rate-dependent part for the Euler angles as dened in [4]. Considering the non-negativity of the volume
fractions, mass conservation, the irreversibility and a maximum value for the stabilized martensite, the
minimization of the Lagrange function results in evolution equations of elasto-plastic type for λ and ρ
β β
λ̇i = devPTi ∀ i ∈ Aλ , ρ̇i = devPTi ∀ i ∈ Aρ (5)
f2 g2
with the thermodynamic driving force
P P T = −∂Ψ/∂λ P = −∂Ψ/∂ρ, the consistency parameter β and the
active deviator devPTi = PTi − ( k∈Aλ PTk /f 2 + l∈Aρ PTl /g 2 )/(nAλ /f 2 + nAρ /g 2 ) using the sets of
active phases Aλ and Aρ . To calculate the consistency parameter β , a Legendre transformation of the
dissipation function is performed, which directly results in the yield function Φ and the Karush-Kuhn-
Tucker conditions:

1 1
Φ = 2 devA P T · devA P T + 2 devA P T · devA P T − rT1 2
≤ 0, β ≥ 0, βΦ = 0. (6)
f Aλ g Aρ
Using the corresponding driving forces P R = {Pϕ , Pϑ , Pω } = −∂Ψ/∂α, the minimization of L related to
the rate of the Euler angles gives evolution equations of viscous type for α:
1 1 1
ϕ̇ = √ (Pϕ − Pω cos ϑ) ϑ̇ = √ Pϑ
2rR 1 − cos2 ϑ 2rR
1 1
ω̇ = √ 2ϑ
(Pω − Pϕ cos ϑ) . (7)
2rR 1 − cos
Figure 1 shows the results for a cyclically loaded clamping ring. The force/displacement diagram displays
the expected material behavior: an increase of the plateau stresses and an accumulating remaining
8
strain. In addition to that, the austenitic volume fraction at maximum load and the total irreversible
martensite ρ̄ after ten load cycles show as a result of the coupled dissipation function the stabilization
of martensitic variants in the transforming areas. A more detailed description of the material model
including a calibration, its nite element implementation and simulations as well as a comparison with
experimental data is already published in [9] and [10].
Figure 1. Simultaneous formation of irreversible martensite cyclic loaded clamping ring: force/displacement di-
agram (left ), austenitic volume fraction at maximum load (center ), total irreversible martensitic volume
fraction after ten load cycles (right ) [10].
2 Simultaneous plastic evolution
In the second presented model, our in a chronological sense rst modeling approach [8], which couples
phase transformation with plasticity, is extended by the use of the evolving orientation distribution
function [4] instead of the former grain-wise calculation of the microstructural evolutions. Instead of an
irreversible martensitic volume fraction in the previous section, a plastic strain of every phase i given
by εPi is used as an internal variable. The combined dissipation function for the simultaneous processes
includes now a dissipation parameter (or yield limit) which is a function of the absolute plastic strain to
take into account the fatigue eect and the converging character of the material behavior. In Figure 2,
the force/displacement diagram of a cyclically loaded plate with a hole as well as the related austenitic
volume fraction at maximum load in the tenth cycle and the plastic strain after the same load cycle
are presented. The model with simultaneous plastic deformation also shows the experimentally observed
material behavior. The plastic deformations form in the arc-like transformation zone and eect the
decrease of the transformation plateaus. A detailed description and several simulations for the model can
be found in [8] and [6].
Figure 2. Simultaneous plastic evolution cyclic loaded plate with a hole: force/displacement diagram (left ),
austenitic volume fraction at maximum load in tenth load cycle (center ), total plastic strain after tenth
cycle (right ) [6].
3 Storage of the transformation history
Whereby the rst two material models are only able to describe the functional fatigue during mechanical
cycling, the last model also displays the material behavior during thermal cycling. Besides the reversible
volume fraction λ and the set of Euler angles α considering the polycrystalline structure, the transfor-
mation history of the material is stored by use of the transformation history variable ρ. To take into
account the fatigue eect, an additional energy term Ψθρ (θ, ρ) as well as an accumulating inelastic strain
υ(ρ) are used, which take into account the thermal and the transformation history's inuence. First
9
results for the material point level are presented in Figure 3. On the left, the characteristic stress/strain
diagram is presented, which again shows the eect of functional fatigue. On the right, the evolution of
the austenitic volume fraction in a thermal cycling test is shown. The third model is thus also able to
show the decrease of the transformation temperatures during thermal cycling, as it is observed ,e.g., in
[1]. A detailed description of the model as well as its nite element implementation are presented in [6]
and [7].
Figure 3. Storage of the transformation history cyclic tension test: stress/strain diagram (left ); thermal cycling
test: evolution of the austenitic volume fraction (right ) [6, 7].
Acknowledgments
I kindly acknowledge Univ.-Prof. Dr.-Ing. habil. Stefanie Reese for her support and many helpful and
inspiring discussions.
References
[1] J. Burow. Herstellung, Eigenschaften und Mikrostruktur von ultrafeinkörnigen NiTi-Formgedächtnis-
legierungen. Ruhr-Universität Bochum, Diss., 2010.
[2] C. Carstensen, K. Hackl, and A. Mielke. Nonconvex potentials and microstructures in nite
strain plasticity. Proceedings of the Royal Society of London. Series A: Mathematical, Physical and
Engineering Sciences 458.2018 (2002), pp. 299317.
[3] G. Eggeler, E. Hornbogen, A. Yawny, A. Heckmann, and M. F.-X. Wagner. Structural and functional
fatigue of NiTi shape memory alloys. Materials Science and Engineering: A 378.1 (2004), pp. 2433.
[4] P. Junker. A novel approach to representative orientation distribution functions for modeling and
simulation of polycrystalline shape memory alloys. International Journal for Numerical Methods in
Engineering 98.11 (2014), pp. 799818.
[5] P. Krooÿ, T. Niendorf, P. M. Kadletz, C. Somsen, M. J. Gutmann, Y. I. Chumlyakov, W. W.
Schmahl, G. Eggeler, and H. J. Maier. Functional Fatigue and TensionCompression Asymmetry
in [001]-Oriented Co49Ni21Ga30 High-Temperature Shape Memory Alloy Single Crystals. Shape
Memory and Superelasticity 1.1 (2015), pp. 617.
[6] J. Waimann. Variationelle Modellierung irreversibler Eekte in polykristallinen Formgedächtnislegierun-
gen. Ruhr-Universität Bochum, Diss., 2018.
[7] J. Waimann and P. Junker. Modeling the inuence of thermal and mechanical cycling on polycrys-
talline shape memory alloys (2019 in preparation).
[8] J. Waimann, P. Junker, and K. Hackl. A coupled dissipation functional for modeling the functional
fatigue in polycrystalline shape memory alloys. European Journal of Mechanics-A/Solids 55 (2016),
pp. 110121.
[9] J. Waimann, P. Junker, and K. Hackl. Modeling the Cyclic Behavior of Shape Memory Alloys.
Shape Memory and Superelasticity (2017), pp. 115.
[10] J. Waimann, K. Hackl, and P. Junker. Variational modeling and nite-element simulation of func-
tional fatigue in polycrystalline shape memory alloys. Journal of Optimization Theory and Appli-
cations (2019), pp. 127.
10
MS 01: Applications and Computational Methods in Metamaterials

Connection between micro- and macroscales

in generalized mechanics
∗ 1
Bilen Emek Abali
1 Technische Universität Berlin, Institute of Mechanics MS 2, Einsteinufer 5, 10587 Berlin, Germany
Abstract. Inner structure of a material starts affecting the material response at the macroscale (length scale
of the geometry), as the geometric dimensions decrease. Miniaturization in nowadays technology obliges the
theory to incorporate the inner structure. Examples for these so-called metamaterials are porous structures like
electrodes in batteries, 3D printed structures with varying infill ratio, foam like materials with exotic behavior.
Such an extension is called generalized mechanics and various theories have been emerged, among others, strain
gradient, micropolar, micromorphic theories are well known and established. Mainly two additional difficulties
arise, the extension of stored energy and kinetic energy. This work highlights the possible connection between
microscale (length scale of the inner structure) and macroscale leading to a better comprehension of upscaling
from the microscale to the macroscale. Additional parameters emerge and their interpretation is possible by
using homogenization techniques. Parameter identification is possible by a pure computational procedure of
solving at micro- and macroscales. Specifically for the strain gradient theory, we demonstrate the results of a
possible approach.
Introduction
Especially with additive manufacturing possibilities like in fused deposition modeling (FDM) mostly
known from so-called 3D printers, an inner substructure is possible to implement in a material. Also in
foam like or porous materials, where the high ratio of surface is aimed for as in the case of electrodes
in a battery, an inner substructure is present within the material. We know that in crystallography, as
studied in [24, 28], an inner substructure called a lattice in the atomic length scale, microscale, cause
anisotropy in the micrometer length scale, macroscale. At the microscale the atomic interaction is often
modeled by two body interaction scalar potentials and one of existing upscaling approaches can be
used to homogenize the lattice structure, see for example the approach based on a statistical method as
introduced in [12]. An analogous attempt is used for composite materials or also materials with inclusions.
The microscale considers a matrix embedding fibers or voids being a heterogeneous material and it leads
to a homogeneous material at the macroscale by using approaches as in [10, 11, 15, 19, 20, 26, 29, 30].
Starting with works of [8, 9, 17], a generalized mechanics is used to consider not only first but also second
gradients in displacement at the macroscale such that the energy definition is discussed heavily, among
others, see [1, 7, 13, 22, 23, 27]. An upscaling procedure has been proposed for generalized mechanics as
well, we refer to [4, 5, 16, 21, 25].
Strain gradient theory and a computational comparison verifying its necessity
We briefly explain and then apply the method introduced in [3] for determining parameters in the case
of a generalized mechanics for a porous material. Then we demonstrate the difference between solutions
with strain gradient and without strain gradient computations. Consider a microscale with a distinct
geometric model as seen in Fig. 1 providing a detailed internal substructure at the length scale of the
microscale. For an upscaling this structure to a macroscale, it is beneficial to use a homogenization
approach. Nevertheless, the first assumption for a homogenization method relies on the descriptions
defined at micro- and macroscales. Often, the same material model is used at both scales; however, as
analyzed in detail by [16], this fact is very questionable. Therefore, we start with an assertion that the
∗ Corresponding author: bilenemek@abali.org ORCID iD: 0000-0002-8735-6071
11
Figure 1. Microscale description of the structure, B, with two mate-

rials indicated by different colors. If embedding material,
matrix, is softer, then the round inclusions provide stiff-
ness. On the contrary, matrix can be stiffer such that in
the limiting case the inclusions can be seen as voids and
the structure represents a porous foam.
micro- and macroscale material models are in general different. Consider the case that we use a first order
theory, the classical case in elasticity, leading to the stored energy density, w, and stress, σ, as follows
∂w
w = εij Cijkl εkl , σij = = Cijkl εkl , (1)
∂εij
where and henceforth we use the summation convention over repeated indices. Stiffness tensor, Cijkl ,
denotes the material response, and we may use a linear strain measure:
1
εij = u(i,j) = (ui,j + uj,i ) , (2)
2
with the partial space derivative denoted by comma. This energy is defined at the microscale, for example
with two isotropic materials given by two parameters, λ, µ, as follows:
Cijkl = λδij δkl + µδik δjl + µδil δjk . (3)
At the macroscale, we consider that two materials are homogenized to one single “new” material. The
same energy definition might be used with different parameters, or in a more general sense, we can try to
utilize an energy including strain gradients as well, considering that it be isotropic and centrosymmetric
w̄ = εij C̄ijkl εkl + εij,k D̄ijklmn εlm,n + εij Ḡijklm εkl,m . (4)
The energy densities are not equal, w 6= w̄. In the case of an isotropic and centro-symmetric material,
they read
C̄ijkl = c1 δij δkl + c2 (δik δjl + δil δjk ),
D̄ijklmn = c3 (δij δkl δmn + δin δjk δlm + δij δkm δln + δik δjn δlm )+
+ c4 δij δkn δml +
+ c5 (δik δjl δmn + δim δjk δln + δik δjm δln + δil δjk δmn )+ (5)
+ c6 (δil δjm δkn + δim δjl δkn )+
+ c7 (δil δjn δmk + δim δjn δlk + δin δjl δkm + δin δjm δkl ) ,
Ḡijklm = 0 ,
The additional parameter tensor, D̄ijklmn , vanishes in the case that a linear elastic material model at
the macroscale Ris adequate.
R Under the assumption that the energy within the structure, B, is identical
at both scales, B w dv = B w̄ dv, it is possible to start with known values of λ, µ at the microscale and
determine values at the macroscale, c1 , c2 , . . . c7 . As an example we consider a porous material with
Young’s modulus, E1 , Poisson’s ratio, ν1 , possessing voids modeled with, E2 and ν2 leading to
Eν E
λ= , µ= . (6)
(1 + ν)(1 − 2ν) 2(1 + ν)
We refer to [3] for details of the used algorithm determining the values compiled in Table 1. The
parameters c× in Table 1 are in the unit of force, hence, it is difficult to justify their significance. Therefore,
we use a computational comparison of a beam bending experiment of this homogenized material. As
experimentally proven in [6, 14, 18], a slender beam gets stiffer than calculated by using analytical
methods indicating this so-called “size effect,” where the strain gradient effect is of importance. We refer
to [2] for details and perform simulations for two cases: one with the obtained c× parameters and another
one by setting them zero. The geometry is a simple beam of length ` with a squared cross sectional area
of width `/30. Its length is reduced by preserving the ratio of length to width the same. Technically, we
perform four computations for each case and compare the results in Table 2.
12
Table 1. Parameters obtained by the computational homogenization method based on [3] for the porous structure
seen in Fig. 1.
E1 E1 ν1 = ν2 c1 c2 c3 c4 c5 c6 c7
1 GPa 1 MPa 0.3 420 MPa 320 MPa 0.20 N 0.05 N -0.18 N -0.02 N -0.02 N
Table 2. Computational comparison for different beam lengths, `, simulating the two cases: case 1 is with strain
gradient parameters from Table 1 and case 2 is for vanishing c3 = c4 = c5 = c6 = c7 = 0
case ` tip deflection difference
1 10 mm 0.197 mm -
2 10 mm 0.197 mm 0%
1 1 mm 0.182 × 10−1 mm -
2 1 mm 0.197 × 10−1 mm 8%
1 0.1 mm 0.188 × 10−2 mm -
2 0.1 mm 0.197 × 10−2 mm 5%
1 0.01 mm 0.162 × 10−3 mm -
2 0.01 mm 0.197 × 10−3 mm 18%
Conclusion
For technical applications getting smaller in size as a part of miniaturization, strain gradient theory
becomes important. In this short note, we have briefly shown a possible consideration as well as physical
interpretation of the strain gradient theory, identification of the newly introduced, as well as testing their
significance by simulating the size effect dominating the material response as the dimensions reduce.
Acknowledgments
B. E. Abali’s work is partly funded by a grant from the Daimler and Benz Foundation.
References
[1] B. E. Abali. Revealing the physical insight of a length-scale parameter in metamaterials by ex-
ploiting the variational formulation. Continuum Mechanics and Thermodynamics (2018), pp. 1–
10.
[2] B. E. Abali, W. H. Müller, and F. dell’Isola. Theory and computation of higher gradient elasticity
theories based on action principles. Archive of Applied Mechanics 87.9 (2017), pp. 1495–1510.
[3] B. E. Abali, H. Yang, and P. Papadopoulos. A Computational Approach for Determination of
Parameters in Generalized Mechanics. In: Higher Gradient Materials and Related Generalized
Continua. Ed. by H. Altenbach, W. H. Müller, and B. E. Abali. Springer Nature, Singapore, 2019.
Chap. 1, pp. 1–17.
[4] H. Abdoul-Anziz and P. Seppecher. Strain gradient and generalized continua obtained by homoge-
nizing frame lattices. Mathematics and mechanics of complex systems 6.3 (2018), pp. 213–250.
[5] D. Bigoni and W. J. Drugan. Analytical derivation of Cosserat moduli via homogenization of
heterogeneous elastic materials. Journal of Applied Mechanics 74.4 (2007), pp. 741–753.
[6] C. Q. Chen, Y. T. Pei, and J. Th. M. De Hosson. Effects of size on the mechanical response
of metallic glasses investigated through in situ TEM bending and compression experiments. Acta
Materialia 58.1 (2010), pp. 189–200.
[7] V. A. Eremeyev, L. P. Lebedev, and H. Altenbach. Foundations of Micropolar Mechanics. Springer
Science & Business Media, 2012.
[8] A. C. Eringen. Mechanics of Micromorphic Continua. In: Mechanics of Generalized Continua. Ed.
by E. Kröner. Berlin: Springer-Verlag, 1968, pp. 18–35.
[9] A. C. Eringen and E. S. Suhubi. Nonlinear Theory of Simple Micro-elastic Solids. International
Journal of Engineering Science 2 (1964), pp. 189–203.
13
[10] Z Hashin and S Shtrikman. On some variational principles in anisotropic and nonhomogeneous
elasticity. Journal of the Mechanics and Physics of Solids 10.4 (1962), pp. 335–342.
[11] R. Hill. On Constitutive Macro-variables for Heterogeneous Solids at Finite Strain. Proc. R. Soc.
Lond. A 326 (1972), pp. 131–147.
[12] J. H. Irving and J. G. Kirkwood. The Statistical Mechanical Theory of Transport Processes. IV.
The Equations of Hydrodynamics. The Journal of Chemical Physics 18 (1950), pp. 817–829.
[13] E. A. Ivanova and E. N. Vilchevskaya. Micropolar continuum in spatial description. Continuum
Mechanics and Thermodynamics 28.6 (2016), pp. 1759–1780.
[14] D. Lam, F. Yang, A. C. M. Chong, J. Wang, and P. Tong. Experiments and theory in strain gradient
elasticity. Journal of the Mechanics and Physics of Solids 51.8 (2003), pp. 1477–1508.
[15] V. M. Levin. Determination of composite material elastic and thermoelastic constants. Mechanics
of Solids 11.6 (1976), pp. 119–126.
[16] K. K. Mandadapu, B. E. Abali, and P. Papadopoulos. On the Polar Nature and Invariance Proper-
ties of a Thermomechanical Theory for Continuum-on-Continuum Homogenization. arXiv preprint
arXiv:1808.02540 (2018).
[17] R. D. Mindlin. Micro-structure in linear elasticity. Archive for Rational Mechanics and Analysis
16.1 (1964), pp. 51–78.
[18] T. Namazu, Y. Isono, and T. Tanaka. Evaluation of size effect on mechanical properties of single
crystal silicon by nanoscale bending test using AFM. Journal of Microelectromechanical Systems
9.4 (2000), pp. 450–459.
[19] L. Nazarenko, H. Stolarski, L. Khoroshun, and H. Altenbach. Effective thermo-elastic properties of
random composites with orthotropic components and aligned ellipsoidal inhomogeneities. Interna-
tional Journal of Solids and Structures 136 (2018), pp. 220–240.
[20] S. Nemat-Nasser and M. Hori. Micromechanics: Overall Properties of Heterogeneous Materials.
Vol. 37. Elsevier, 2013.
[21] C. Pideri and P. Seppecher. A second gradient material resulting from the homogenization of
an heterogeneous linear elastic medium. Continuum Mechanics and Thermodynamics 9.5 (1997),
pp. 241–257.
[22] C. Polizzotto. A second strain gradient elasticity theory with second velocity gradient inertia–Part
I: Constitutive equations and quasi-static behavior. International Journal of Solids and Structures
50.24 (2013), pp. 3749–3765.
[23] C. Polizzotto. A second strain gradient elasticity theory with second velocity gradient inertia–Part
II: Dynamic behavior. International Journal of Solids and Structures 50.24 (2013), pp. 3766–3777.
[24] A. Reuss. Berechnung der Fließgrenze von Mischkristallen auf grund der Plastizitätsbedingung für
Einkristalle. ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift für Angewandte
Mathematik und Mechanik 9.1 (1929), pp. 49–58.
[25] P. Seppecher, J.-J. Alibert, and F. dell’Isola. Linear elastic trusses leading to continua with exotic
mechanical interactions. In: Journal of Physics: Conference Series. Vol. 319. IOP Publishing. 2011,
p. 012018.
[26] B. Shafiro and M. Kachanov. Anisotropic effective conductivity of materials with nonrandomly
oriented inclusions of diverse ellipsoidal shapes. Journal of applied physics 87.12 (2000), pp. 8561–
8569.
[27] P. Steinmann. A micropolar theory of finite deformation and finite rotation multiplicative elasto-
plasticity. International Journal of Solids and Structures 31.8 (1994), pp. 1063–1084.
[28] W. Voigt. Über die Beziehung zwischen den beiden Elasticitätsconstanten isotroper Körper. An-
nalen der Physik 274.12 (1889), pp. 573–587.
[29] T. I. Zohdi. Homogenization methods and multiscale modeling. Encyclopedia of Computational
Mechanics Second Edition (2017), pp. 1–24.
[30] T. I. Zohdi and P. Wriggers. An Introduction to Computational Micromechanics. Springer Science
& Business Media, 2008.
14

On the modeling of microstructure influence

∗
Arkadi Berezovski
Tallinn University of Technology, School of Science, Tallinn, Estonia
Abstract. The mathematical structure of micromorphic elasticity is analyzed from the point of view of the
internal variables theory. It is demonstrated how dispersion characteristics of governing equations depend
on the choice of sign for material parameters. This choice provides the explanation of the results of direct
numerical simulations of a pulse propagation in an inhomogeneous one-dimensional bar.
Introduction
There exist various methods to take into account the influence of microstructure of materials on their
global behavior, e.g., distinct homogenization approaches [7, 8], strain gradient [5], micropolar, and
micromorphic theories [4]. It is not easy task to choose an appropriate description due to a lack of values
of material parameters. In the paper, we analyze the mathematical structure of micromorphic elasticity
and its relation to the wave propagation features. For simplicity, we constrain ourself by one-dimensional
setting.
1 Free energy
It is well known that linear elasticity obeys the Hooke law

σ = ρc2 ux , (1)
where σ is the stress, ρ is the density, u is the displacement, c is the elastic wave speed. The Hooke
law corresponds to the free energy density in the form
ρc2 2
W = u . (2)
2 x
Certainly, there is no any influence of a microstructure. An extension of the free energy density is needed
to include such an effect. The first possibility of such extansion is delivered by the inclusion of the strain
gradient into the elastic model
ρc2 1
W = (ux )2 + Aux uxx + B(uxx )2 . (3)
2 2
It should be noted that the strain gradient can be considered as an internal variable ϕ = A/B uxx [2],
and the free energy density is then represented in the form
ρc2 2 1
W = ux + Aϕux + Bϕ2 . (4)
2 2
Of course, this is only the first step in the generalization of the free energy density. To go further, we
consider the internal variable as a new unknown field and include space and time rates of the internal
variable keeping the quadratic dependence of the free energy density
ρc2 2 1 1 1
W = ux + Aϕux + Bϕ2 + Cϕ2x + M ϕ2t . (5)
2 2 2 2
∗ Corresponding author: arkadi.berezovski@cs.ioc.ee
15
It should be noted that the last relation corresponds to the Mindlin microelasticity in one dimension if
B = −2A [1]
ρc2 2 1 1
W = u + Aϕux − Aϕ2 + Cϕ2x + M ϕ2t . (6)
2 x 2 2
2 Equations of motion
In the linear elasticity the equation of motion is the well known wave equation
utt = c2 uxx . (7)
The equation of motion for the strain gradient model is extended to
A2
ρutt = ρc2 uxx − uxxxx , (8)
B
and in the case of internal variables we have the system of equations
ρutt = ρc2 uxx + Aϕx , (9)

M ϕtt = Cϕxx − Aux − Bϕ, (10)
As previously, the choice B = −2A leads to Mindlin microelasticity (and to relaxed micromorphic theory
[6]) in one-dimensional setting.
2.1 Dimensionless variables
For further analysis we introduce dimensionless variables

u x c0 t U0
U= , X= , T = , ϕ= Φ, (11)
U0 L L L
where U0 and L are certain constants (e.g., intensity and wavelength of the initial excitation), and c0 is
the characteristic velocity.
Introducing dimensionless parameters
A B C I
a= . b= , c= , m= (12)
ρc20 ρc20 L2 ρc20 L2 ρ
we have for the microstructure model with internal variables
UT T = UXX + aΦX , (13)

mΦT T = cΦXX − aUX − bΦ. (14)
Again, the choice b = −2a leads to the Mindlin microelasticity.
3 Dispersion relation
In order to study dispersive effects, we derive the dispersion relations by assuming the solutions in the
form of harmonic waves
U (x, t) = Û ei(kx−ωt) , Φ(x, t) = Φ̂ei(kx−ωt) , (15)
where k is the wavenumber, ω is the frequency, and i2 = −1. Substituting relations (15) into Eqs. (13)
and (14) we get
−Û ω 2 = −Û k 2 + aΦ̂(ik), (16)

2 2
−mΦ̂ω = −cΦ̂k − aÛ (ik) − bΦ̂. (17)
16
This can be represented in the matrix form

2
k − ω2 −iak Û
= 0, (18)
iak ck 2 − mω 2 + b Φ̂
In order to get nontrivial solutions the determinant of this matrix must vanish, which results in
1 h 2 p i
ω2 − k2 = (ck − mk 2 + b) ± (ck 2 − mk 2 + b)2 + 4ma2 k 2 . (19)
2m
Specifying the values of dimensionless parameters as m = 1, b = 1, a = 0.5 c = 3, we have for the
microstructure model with internal variables
" r #
ω2 1 1 1 2 1
−1= (2 + 2 ) ± (2 + 2 ) + 2 . (20)
k2 2 k k k
We are interesting in the behavior of short waves, because the long wave approximation is well studied.
For short waves values of the wave number are high and, therefore, ε = 1/k 2 is a small parameter. Then
the dispersion relation is represented in the form
2 1h p i
vph =1+ (2 + ε) ± (2 + ε)2 + ε (21)
2
with phase velocity vph , and the corresponding dispersion curves are shown in Fig. 1. These curves
represent the Mindlin microelasticity if a = −0.5.
3
ω2/κ2
0
0 0.2 0.4 0.6 0.8 1
ε
Figure 1. Dispersion curves corresponding to Eq. (21).
As it is demonstrated by direct numerical simulations [1], the Mindlin microelasticity model (as well
as its extension in terms of internal variables) is unable to reproduce a pulse propagation behavior in
an inhomogeneous bar if the size of inhomogeneities is comparable with the wavelength. However, the
modification of the signs of material parameters resulting in equations of motion
UT T = UXX + aΦX , (22)

mΦT T = cΦXX + aUX + bΦ, (23)
provides much more reliable results [3]. The corresponding dispersion relation
2 1h p i
vph =1+ (2 − ε) ± (2 − ε)2 − ε (24)
2
is illustrated in Fig. 2. As one can see, the dispersion curves are substantially different from those in the
previous case.
17
2.5
ω2/κ2
2
1.5
1
0 0.2 0.4 0.6 0.8 1
ε
Figure 2. Dispersion curves corresponding to Eq. (24).
4 Conclusions
Analysis of the structure of micromorphic elasticity theory shows that its representation in terms of
internal variables includes both the Mindlin microelasticity and relaxed micromorphic model (at least
in one dimension). For short wave lengths, all these models are unable to reproduce the results direct
numerical simulations of a pulse propagation in an inhomogeneous bar. The change of signs of some
material parameters in the microstructure model provides more reliable results consistent with numerical
results.
Acknowledgments
The work is supported by the Centre of Excellence for Nonlinear Dynamic Behaviour of Advanced Ma-
terials in Engineering CZ.02.1.01/0.0/0.0/15 003/0000493 (Excellent Research Teams) in the framework
of Operational Programme Research, Development and Education, by the Grant project with Nos. 19-
04956S of the Czech Science Foundation (CSF) within institutional support RVO:61388998, and by
Estonian Research Council under Institutional Research Funding IUT33-24.
References
[1] Arkadi Berezovski. On the Mindlin microelasticity in one dimension. Mechanics Research Commu-
nications 77 (2016), pp. 60–64.
[2] Arkadi Berezovski. Internal variables associated with microstructures in solids. Mechanics Research
Communications 93 (2018), pp. 30–34.
[3] Jüri Engelbrecht, Arkadi Berezovski, and Mihhail Berezovski. Deformation waves in microstructured
materials: theory and numerics. In: IUTAM Symposium on Recent Advances of Acoustic Waves in
Solids. Ed. by Tsung-Tsong Wu and Chien-Ching Ma. Springer. 2010, pp. 19–29.
[4] Samuel Forest. Micromorphic media. In: Generalized Continua from the Theory to Engineering
Applications. Ed. by Holm Altenbach and Victor A. Eremeyev. Springer, 2013, pp. 249–300.
[5] Raymond David Mindlin and NN Eshel. On first strain-gradient theories in linear elasticity. Inter-
national Journal of Solids and Structures 4 (1968), pp. 109–124.
[6] Patrizio Neff, Ionel-Dumitrel Ghiba, Angela Madeo, Luca Placidi, and Giuseppe Rosi. A unifying
perspective: the relaxed linear micromorphic continuum. Continuum Mechanics and Thermodynam-
ics 26 (2014), pp. 639–681.
[7] Siavouche Nemat-Nasser and Muneo Hori. Micromechanics: Overall Properties of Heterogeneous
Materials. Elsevier, 2013.
18

A variational-based energy-momentum time integration

of metamaterials in non-isothermal rotordynamical sys-
tems
1
Michael Groß ∗
, Julian Dietzsch1 and Chris Röbiger1
1 TU Chemnitz, Professorship of applied mechanics and dynamics, Reichenhainer Straße 70, D-09126 Chemnitz, Germany.
Abstract. Textile manufacturing techniques as three-dimensional weaving of carbon yarns or the tailored fiber
placement based on fixing carbon rovings on a base material are increasingly used for parts in rotordynamical
systems. Examples are blades in turbine fans and pump rotors, disc brake rotors and shafts. These metama-
terials allow for lower rotational masses while maintaining stiffness and improving heat conduction properties.
This leads in rotordynamical systems to higher possible rotational speeds. Thereby, the direction-dependent
heat conduction allows for a more efficient cooling in high temperature environments. A continuum formulation
of such metamaterials has to take into account the bending stiffness of the yarns or rovings, because carbon
yarns and rovings are made of thousends of carbon fibers. A standard Cauchy-continuum with a free energy
based on structural tensors are not able to describe this fiber bending stiffness, because fibers are assumed as
infinitely thin strings. This results in a three-point bending test of metamaterials simulated by finite elements
to wrong curvature effects, in comparison to experimental data. The bending stiffness of the yarns or rovings
can be taken into account by tensor invariants based on the second gradient of the deformation mapping,
which leads to a generalized continuum formulation. In this paper, we consider a mixed formulation of a
generalized continuum, in which the deformation gradient and the first Piola- Kirchhoff stress tensor as well as
the continuum rotation vector, the angular velocity vector and the rotational momentum vector are introduced
as independent fields. The corresponding weak forms arise from a principle of virtual power, which allows for
energy-momentum time integrations of the considered rotordynamical systems. In this principle, the kinetic
energy includes the translational energy of a standard Cauchy continuum based on the linear momentum, and
the additional rotational energy with a micro-inertia tensor of a generalized continuum. This leads to two
additional balance laws in comparison to a standard Cauchy continuum: the balance law of rotational momen-
tum and the balance law of rotational energy. The exact preservation of these two balance laws along with
the preservation of each balance law of a standard Cauchy continuum allows the derived energy-momentum
scheme. We demonstrate the behavior of this physically-consistent simulation method by using dynamic sim-
ulations of a turbine rotor, a disc brake rotor and a rotating shaft subject to thermal and mechanical loads.
Introduction
In this paper, we show our first step, a transversely isotropic first-gradient material law in the variational
setting of a generalized continuum. This assumption fits to the tailored fiber placement technique.
We consider non-isothermal deformations of a transversally isotropic continuum moving in the Euclidean
space Rndim with the constant ambient temperature Θ∞ . The initial configuration B0 = B0M ∪ B0F of the
fiber-reinforced continuum body B (see Fig. 1) is defined as homogenization of the set B0M of the matrix
material with the set B0F of the fiber bundels or rovings. In this way, we consider an imaginary fiber at any
point X ∈ B0 , which is directed along the normalized fiber vector a0 with ka0 k = 1. The corresponding
stretched fiber vector a = F a0 in the current configuration Bt is given by the deformation gradient
F := Grad[ϕ] of the deformation mapping ϕ : B0 → Bt . The operator Grad[•] indicates the partial
derivative with respect to the material point X ∈ B0 . The deformation gradient F F in fiber direction
takes the form F F := a ⊗ a0 = F a0 ⊗ a0 = F A0 , with A0 := a0 ⊗ a0 = F −1 [a ⊗ a] F −t =: F −1 A F −t .
We denote by the superscript letter t the transposition of a second order tensor. The tensors A0 and A
denote the material and spatial structural tensors, respectively. The deformation tensor
t
C F := F tF F F = [F a0 ⊗ a0 ] [F a0 ⊗ a0 ] = [a0 ⊗ a0 ] C [a0 ⊗ a0 ] = A0 C A0 = [C : A0 ] A0 (1)
∗ Corresponding author: michael.gross@mb.tu-chemnitz.de
19
N Q̄ Θ∞
T̄
∂B0 ϕ ∂Bt
Θ=θ◦ϕ θ
C v = F tv F v F̃ 6= F = ∇ϕ Tx Bt
t0
tt
t
a0 C̃ 6= C = F̃ F̃ a
X x
A0 At
TX B0
B0 B Fv Fe Bt
V
Figure 1. The initial and current configuration of a fiber-reinforced continuum with mechanical and thermal
boundary as well as volume loads.
of the fibers is related to the right Cauchy-Green tensor C := F t F by a projection on A0 . The deforma-
tion of the isotropic matrix material complies with the multiplicative split F = F e F v of F in the viscous
deformation gradient F v and the elastic deformation gradient F e (see Fig. 1). Here, the tangent space
V denotes the so-called ‘intermediate configuration’. Describing only isotropic viscous behaviour of the
matrix material, we introduce the viscous right Cauchy-Green tensor C v := F tv F v as internal variable.
We consider the total free energy density Ψ (X) := Ψ ela (X) + Ψ the (X) + ΨM vis
(X) with respect to the
initial configuration at any point X ∈ B0 . The elastic free energy Ψ is given by the function
ela
Ψ̂ ela (C, CV , CF ; A0 ) = Ψ̂M

iso C
(I1 , I2C , I3C , I4C ; A0 ) + Ψ̂M
vol
(CV ) + Ψ̂Fela (CF ) (2)
where the semicolon in the argument separates the parameter tensor A0 acting at any X ∈ B0 from the
variables C, CV and CF . Here, we denote by
1 C 2
I1C := C : I I2C := (I1 ) − C 2 : I I3C := det C I4C := C : A0 = C F : I (3)
2
the isotropic tensor invariants of the tensor pair (C, A0 ), where the tensor I designates the second-order
identity tensor. The invariant I4C = C : A0 ≡ a0 · C a0 = a · a = kak2 designates the squared fiber
1
stretch λ2F , and the tensor invariant CV := [det C] ndim the volumetric invariant of the right Cauchy-Green
tensor. The thermal free energy Ψ the = Ψ̂ the (Θ, CV , CF ) is defined by the function
q q
cap
Ψ̂ the := Ψ̂M (Θ)− 2 C̃V2−ndim βM [Θ−Θ∞ ] DΨ̂M vol
(C̃V )+Ψ̂Fcap (Θ)− 2 C̃F2−1 βF [Θ − Θ∞ ] DΨ̂Fela (C̃F ) (4)
cap
where the heat capacity function Ψ̂ cap is separated in two parts Ψ̂M and Ψ̂Fcap according to the volume
fraction of matrix material and fiber material, respectively. The notation D(•) here denotes the Fréchet
derivative of a volume density with respect to its argument. The material parameter βM and βF are the
coefficients of linear thermal expansion of the matrix material and fiber material, respectively. The field
Θ(X) denotes the absolute temperature Θ = θ ◦ ϕ of the current temperature θ(x) at a point x ∈ Bt .
The viscoelastic free energy ΨM vis
of the matrix material is defined by the function Ψ̂M (I1 , I2 , I3 ) with
vis Λ Λ Λ
1 Λ 2
I1Λ := Λ : I I2Λ := (I1 ) − Λ2 : I I3Λ := det Λ (5)
2
as invariants pertaining to the product tensor Λ := CC −1 v , which depends explicitly on the viscous
internal variable C v . These invariants coincide with the invariants of the deformation tensor C e = F te F e .
In view of the introduction of a second-gradient free energy in the next steps of the project, we now intro-
duce an independent deformation gradient F̃ and right Cauchy-Green tensor C̃, which is, in the discrete
setting, different from F and C, but related by the mixed formulation in Section 1. The deformation of
the fibers is then described by the projection of the independent deformation gradient F̃ on the structural
tensor A0 . The continuum rotation vector α, also introduced in Section 1, is likewise defined upon the
mixed field F̃ . Furthermore, we introduce independent fields C̃F and C̃V for fiber and volume dilatation,
because the free energy function Ψ̂Fela (C̃F ) depends explicity on the squared fiber stretch C̃F , and the free
energy functions Ψ̂M vol
(C̃V ) and Ψ̂ the (Θ, C̃V , C̃F ) also depend explicitly on the volumetric constant C̃V .
20
1 Variational formulation
The variational formulation of the considered generalized Cauchy continuum is based on the principle of
virtual power, in which the total energy balance law is considered in the functional form
Ḣ(ϕ̇, v̇, ṗ, α̇, ω̇, π̇, Ċ v , Θ̇, η̇, F̃˙ , C̃,
˙ C̃˙ , C̃˙ , Θ̃, P , τ t , S, S , S , R, h, λ, Z, ω̂) = 0
V F skw V F (6)
where Ḣ = Ṫ + Ṫ + Π̇ + Π̇ denotes the time rate of the considered total energy H, consisting of
tra rot ext int
the translational and rotational kinetic energies, respectively, as well as the external and internal potential
energy of the continuum body. The well-known translational kinetic power of a continuum is given by
Z Z Z
tra
Ṫ (ϕ̇, v̇, ṗ) := [ρ0 I v − p] · v̇ dV − ṗ · [v − ϕ̇]dV + p · ϕ̈ dV (7)
B0 B0 B0
with respect to the homogenized Lagrangian mass density ρ0 of the body in B0 . The Lagrangian velocity
field of the body is indicated by v, and the conjugated Lagrangian momentum field is denoted by p. In
order to take into account the micro inertia of the considered heterogeneous material, we introduce
Z Z Z
2
Ṫ rot (α̇, ω̇, π̇) := ρ0 (lF − l02 ) A0 + l02 I ω − π · ω̇ dV − π̇ · [ω − α̇]dV + π · α̈ dV (8)
B0 B0 B0
as rotational kinetic power of the continuum. Here, we denode by lF and l0 length scale parameters
corresponding to the fiber bundels and the considered representative volume element, respectively. The
Lagrangian vector field ω denotes the Lagrangian angular velocity field, and π is the conjugated La-
grangian rotational momentum field. The Lagrangian vector field α describes the continuum rotation,
which is defined in the internal potential power functional
Z (" # " # ) Z h i
int 1 ∂ Ψ̂M vol ˙ ∂ Ψ̂M ˙
Π̇ := 2 + SV A + SF A0 − S : C̃ + 2 − SV : C̃V dV − η̇ Θ̃ − Θ dV
2 B0 ∂ C̃ ∂ C̃V B0
Z " # Z ( ) Z
1 ∂ Ψ̂F h i ∂ Ψ̂M ∂Ψ
+ 2 − SF : C̃˙ F dV + F̃ S − P : F̃˙ + : Ċ v dV + η+ Θ̇ dV
2 B0 ∂ C̃F B0 ∂C v B0 ∂Θ
Z Z
1
: F̃˙ F̃ + α̇ dV
−1
+ P : Grad[ϕ̇] dV + τ tskw : ·
B0 B0 2
(9)
by using the third-order permutation tensor . The free energy function Ψ̂M (C̃, C v , C̃V , Θ) belongs to the
matrix material, and Ψ̂F represents the fiber free energy. We also introduce independent (mixed) stress
fields S, SV and SF , energy-conjugated to the right Cauchy-Green tensor C̃, the volumetric invariant
C̃V and the squared fiber stretch C̃F , respectively. Beside the absolute temperature field Θ, the thermal
behaviour of the continuum is described by further mixed fields. We consider the entropy density field η
as well as the assumed temperature field Θ̃, which, in fact, approximate the time derivative of the thermal
displacement field. Further, we denote by P the first Piola-Kirchhoff stress tensor, and the Lagrange
multiplier τ tskw designates the skew-symmetric part of the Kirchhoff stress. Since we simulate continuum
bodies subject to volume and boundary loads as well as apply a variational-based energy-momentum
scheme, we have to consider the external power functional
Z Z Z
ext Θ̃
Π̇ := − ρ0 B · ϕ̇ dV − T̄ · ϕ̇ dA + Q̄ dA
B ∂ B ∂ B Θ
Z 0 Z T 0 ZQ 0
1 Θ̃
+ Grad[Θ̃] · Q dV + Dcdu + Dint dV + Ċ v : Σ v dV
Θ Θ
ZB0 h i Z B0
h i Z B0
+ λ Θ̃ − Θ∞ dA − h Θ̇ − Θ̄˙ dA − ˙ dA
R · [ϕ̇ − ϕ̄] (10)
∂Θ B0 ∂Θ̇ B0 ∂ϕ B0
Z Z Z

− ˙ dA −
Z · [α̇ − ᾱ] ω̂ · : τ tskw dA − W̄ · α̇ dA
∂α B0 ∂α B0 ∂W B0
Z Z Z
1 ˙ dV 1 1
+ S̄ : C̃ + S̄V C̃˙ V dV + S̄F C̃˙ F dV
B0 2 B0 2 B0 2
with Dirichlet and Neumann boundaries. The stationarity condition of the virtual power principle reads
δ∗ Ḣ(ϕ̇, v̇, ṗ, α̇, ω̇, π̇, Ċ v , Θ̇, η̇, F̃˙ , C̃,
˙ C̃˙ , C̃˙ , Θ̃, P , τ t , S, S , S , R, h, λ, Z, ω̂) = 0
V F skw V F (11)
where δ∗ indicates the variation with respect to each tensor in the argument, and leads to all weak forms.
21
2 Numerical example
We simulate a driven rotation of a composite turbine rotor under thermal loads (see Fig. 2). The fibers
in the blades lie in normal direction, and in the shaft in tangential direction. The rotation of the turbine
is introduced by means of a boundary torque W̄ (t) on the frontside and the backside of the shaft. On the
red patches, we load the plates with a periodic boundary heat flux Q̄(t) for cooling, and with a pressure
follower load T̄ (t) leading to a braking rotation. The mesh is derived from Reference [1]. In Fig. 2, we
show the time evolutions of a node on the drive torque boundary. The time evolution ω(t) of the angular
velocity vector reveals oscillations due to the braking pressure. The heat flux leads to an oscillating time
evolution η(t) of the entropy. The z-component of the velocity v(t) reveals also a translational motion.
Figure 2. Time evolutions of a node on the boundary ∂W B0 on the frontside of the shaft.
Acknowledgments
This research is provided by the ’Deutsche Forschungsgemeinschaft’ (DFG) under the grants GR 3297/4-1
and 3297/6-1. This support is gratefully acknowledged.
References
[1] Ludovic Noels, Laurent Stainier, and Jean-Philippe Ponthot. A Consistent Dissipative Time Integra-
tion Scheme for Structural Dynamics: Application to Rotordynamics. In: 45th AIAA/ASME/ASCE/
AHS/ASC Structures, Structural Dynamics & Materials Conference. 2004, p. 1945.
22

Elastic wave propagation in periodical one-dimensional

problems: Accurate nite element modelling
Radek Kolman
∗ 1, 2 3 4 5
Sang Soon Cho , K.C. Park , Arkadi Berezovski , José González , Vít¥zslav Adámek
6
and Pert Hora
7
1 Institute of Thermomechanics, Academy of Sciences of the Czech Republic, v.v.i. Dolejskova 5, Prague 8, Czech Republic
2 Reactor Mechanical Engineering Devision, Korea Atomic Energy Research Institute, 999-111 Daedeok-Daero, Yuseong-gu, Dae-
jeon 305-353, Korea

3 Department of Aerospace Engineering Sciences, University of Colorado, Boulder, CO 80309-429, USA
4 Institute of Cybernetics at Tallinn University of Technology, Akadeemia tee 21, 12618 Tallinn, Estonia
5 Escuela Técnica Superior de Ingeniería, Universidad de Sevilla, Camino de los Descubrimientos s/n, Sevilla 41092, Spain
6 NTIS - New Technologies for the Information Society, Faculty of Applied Sciences, University of West Bohemia, Technická 8,
301 00 Pilsen, Czech Republic

7 Institute of Thermomechanics, v.v.i., The Czech Academy of Sciences, Veleslavínova 71/11, 301 14 Pilsen, Czech Republic
Abstract. In this contribution, we present results of novel explicit nite element approach for accurate mod-
elling of elastic wave propagation in periodical one-dimensional layered bars. Standard approaches in explicit
time integration of the nite element discretized system produce dominant spurious oscillations of stress dis-
tributions. By meaning of the local stepping process, the stress spurious oscillations are eliminated and the
time scheme is suitable for accurate modelling of elastic wave propagation in heterogeneous media using the
nite element method. The presented time scheme is tested in elastic wave propagation in a layered bar with
two dierent materials.
Introduction
Elastic wave propagation in heterogeneous solids is an up-to-date open problem to study due to new
growing and modern technologies for 3D printing of heterogeneous materials. Today it is possible to
use 3D/4D printing technologies for manufacturing of bodies made from several materials as layered,
composite or graded structures. The numerical modelling is necessary because macroscopic properties of
composite materials are strongly inuenced by the properties of their constituents and micro-structures.
Layered periodical materials or metamaterials with auxetic behaviour represent the simplest possible
pattern of composites from the theoretical point of view [3]. Their modelling as periodic structures also
has a rich history, see the book of [2]. Several geometrical problems of elastic wave propagation in layered
media have been studied in book of [7]. Constitutive models of eective properties for such materials are
still under development using either ensemble averaging or homogenization. For short waves, it is worth
to use numerical simulations for the analysis of interaction between layers and waves taking into account
reected and transmitted waves [1].
We present results of novel explicit nite element approach for accurate modelling of elastic wave propa-
gation in periodical one-dimensional layered bars [6]. It is known that the standard approach in explicit
time integration of the nite element discretized system produces spurious oscillations of stress distribu-
tions and the nal stress states are numerically polluted. From that reason, we have derived and applied
the modication of explicit time integration with the local stepping process based on correct local wave
speed and pullback interpolation [5, 8]. With this approach, the numerical dispersion of the nite element
distretization and sensitivity on time step size are suppressed.
∗ Corresponding author: kolman@it.cas.cz ORCID iD: 0000-0002-8167-3535
23
1 Time integration
We compare two explicit time schemes for nite element modelling of wave response with the nite
element method: the central dierence method in time [4] and local time stepping scheme [6]. The
used local time stepping explicit scheme is consisted of two following computational steps based on the
predictor-corrector form:
Step 1: Pull-back integration with local stepping

a) Integration by the central dierence scheme with the local (elemental) critical time step size for each
nite element.
b) Pull-back interpolation of local nodal displacement vector at the time tn+1 .
c) Assembling of local contributions of displacement vector from Step 1b.
Step 2: Push-forward integration with averaging
a) Push-forward integration by the central dierence scheme with the time step size.
b) Averaging of the total displacement vectors at the time tn+1 form Steps 1c and 2a.
c) Evaluation of acceleration and velocity nodal vectors at the time tn+1 .
2 Wave propagation in a layered bar
As a comparative test, we study an problem of elastic wave propagation in a layered bar of the length
L = 2 m with ten uniform layers, for detail see Fig. 1. Mechanical parameters of materials are chosen
as follows: density of materials ρ1 = ρ2 = 1 kg/m3 , Young's moduli E1 = 16 P a and E2 = 1 P a. Thus
wave speeds are given as c1 = 4 m/s and c2 = 1 m/s and acoustic impedance Z1 = 4 kgm−2 s−1 and
Z2 = 1 kgm−2 s−1 . The bar is loaded by the rectangular pulse with stress amplitude σ0 = −1 P a and
with time duration T0 = 0.25 s. The nal time of computations is Tend = 1.2 s. The bar is xed on the
right hand side and the bar is at the rest at the beginning of computations.
Figure 1. Formulation of wave propagation problem on a layered bar.
The stress distributions at the time 1.2 s for the both time schemes are presented in Fig 2. One can see
the better stress distribution for the local time stepping than for the classical central dierence method,
which produces dominant spurious oscillations.
3 Conclusions
Based on the presented numerical test on wave propagation in periodical one-dimensional structurebar,
we shown accuracy of the presented approach in the nite element modelling. The results do not exhibit
stress spurious oscillations against the classical explicit time integration. The local time stepping explicit
scheme is a suitable and robust tool for accurate modelling of elastic wave propagation in one-dimensional
24
Figure 2. Spatial stress distributions in a layered bar from Fig. 1 obtained by the central dierence method
(on the left) and the novel local time stepping scheme (on the right).
heterogeneous cases. In future, we will pay an attention on multi-dimensional problems of elastic wave
propagation tasks in heterogeneous media and solids.
Acknowledgments
The work was supported by the Grant project with No. 19-04956S of the Czech Science Foundation (CSF)
within institutional support RVO: 61388998 and the Centre of Excellence for Nonlinear Dynamic Be-
haviour of Advanced Materials in Engineering CZ.02.1.01/0.0/0.0/15_003/0000493 (Excellent Research
Teams) in the framework of Operational Programme Research, Development and Education.
References
[1] A. Berezovski, J. Engelbrecht, and G.A. Maugin. Numerical Simulation of Waves and Fronts in
Inhomogeneous Solids. World Scientic, 2008.
[2] L. Brillouin. Wave Propagation in Periodic Structures: Electric Filters and Crystal Lattices. McGraw-
Hill Book Company, 1946.
[3] K.F. Gra. Wave Motion in Elastic Solids. Clarendon Press, 1975.
[4] T.J.R Hughes. The Finite Element Method: Linear Static and Dynamic Finite Element Analysis.
Dover Civil and Mechanical Engineering, Dover Publications., 2008.
[5] R. Kolman, S.S. Cho, and K.C. Park. Ecient implementation of an explicit partitioned shear and
longitudinal wave propagation algorithm. International Journal for Numerical Methods in Engineer-
ing 107 (2016), pp. 543579.
[6] R. Kolman, S.S Cho, J.G. Gonzalez, K.C Park, and A. Berezovski. An explicit time scheme with
local time stepping for one-dimensional wave propagation in a bimaterial bar. In conference Engi-
neering Mechanics 2017, University of technology, Institute of Solid Mechanics, Mechatronics and
Biomechanics, 2017, pp. 474477.
[7] A.H. Nayfeh. Wave Propagation in Layered Anisotropic media. North-Holland, 1995.
[8] K.C. Park, S.J.A Lim, and H. Huh. A method for computation of discontinuous wave propagation
in heterogeneous solids: basic algorithm description and application to one-dimensional problems.
International Journal for Numerical Methods in Engineering 91 (2012), pp. 622643.
25
26

C0 continuous finite elements for gradient elasticity at

finite strains
Johannes Riesselmann1 , Jonas Ketteler2 , Mira Schedensack2 and Daniel Balzani ∗ 1
1 Chair of Continuum Mechanics, Ruhr-Universität Bochum, Universitätsstraße 150, 44801 Bochum, Germany
2 Institute for Analysis and Numerics, Westfälische Wilhelms-Universität Münster, Einsteinstraße 62, 48149 Münster, Germany
Abstract. Purely displacement based gradient elasticity formulations require C 1 -continuous finite elements,
for which retaining compatibility with standard software and meshing arbitrary structures are known diffi-
culties. Another approach is to use mixed formulations instead. They have the advantage of a relaxation of
continuity requirement to C 0 and making standard finite element discretizations possible. For small strains
already existing mixed three-field approaches contain both displacement and displacement gradients as so-
lution variables. In this contribution a new formulation is proposed, in which the displacements are not
part of the problem anymore, but only displacement gradients, leading to a reduced number of variables
to be discretized. In numerical examples the proposed formulations are compared to finite strain adapta-
tions of existing mixed three-field approaches and unveil favorable computing time. A further reduction
of variables through a penalty formulation is investigated and the ability to model size effects is shown.
1 Gradient elasticity fundamentals
Gradient elasticity formulations can be used to model specialized materials, whose elastic behavior de-
pends on size relations between the microstructure and specimen size. Through enrichment of the internal
elastic energy W ··= W (F , ∇F ) by the second order gradient of deformation ∇F and additional micro-
size related constitutive parameters, these so called size effects can be captured. Moreover, singularities
at e.g. sharp corners of the modeled specimen, which would appear in classical elasticity formulations
are avoided and corresponding numerical simulations have the advantage of remaining mesh objective.
The theoretical foundation for the small strain gradient-enriched continuum theory dates back to the
works of [5] and [10] and the more recent adaptation of [1]. A historical overview is given in [2] and an
investigation of finite strain models can be found in [8] and [11].
The purely displacement based model problem seeks the minimizer of the potential Π = Π int + Π ext ,
where Π ext is the potential of surface tractions and volume loads and
ˆ
Π int = W (F (∇u), ∇F (∇u)) dV (1.1)
B
is the internal elastic potential. The solution is then sought in the Sobolev space u ∈ H 2 (B) of functions,
whose gradients up to second order are L2 (B) integrable. This leads to the requirement of C 1 -continuous
interpolation functions, for which standard finite elements are not suitable. Isogeometric analysis e.g.
as considered by [8] can satisfy the continuity requirement but has the kown difficulties when modeling
complex structures and boundary conditions. Another approach is to use three-field mixed formulations
instead (c.f. [9] and [12] for the small strain framework and [7] for corresponding large strain formulations).
The idea is to replace ∇u in the gradient-enriched part of W by the new independent variable H ∈ H 1 (B)
and enforce compatibility via the Lagrange multiplier λ ∈ L2 (B) such that the internal potential reads
ˆ
Π int = W (F (∇u), ∇F (H)) + λ : (H − ∇u) dV. (1.2)
B
Consequently, the continuity requirement is relaxed to C 0 , at the cost of having to approximate the three
variables u, H and λ in one BVP. Moreover, in the linear case, inf-sup stability of formulation (1.2) with
∗ Corresponding author: daniel.balzani@rub.de
27
respect to the standard norms is not fulfilled [7] and thus, robustness of the solution procedure in the
nonlinear case is not guaranteed. In the following, an alternative C 0 continuous approach with a reduced
number of solution variables is presented, which for the related biharmonic problem has already proven
to be stable (c.f. [4]).
2 Curl-free mixed two-field approach
In order to reduce the number of degrees of freedom, the displacement u can be decoupled from the
problem in the spirit of [4, 6]. The reformulation of (1.1) now reads
ˆ ˆ
int ext
Π = W (F (H), ∇F (H)) + λ : rot H dV, Π = − ∇g : H dV (2.1)
B B
making H and λ the only solution variables, while retaining the C 0 -continuity requirement. H is sought
in the standard solution space H ∈ H 1 (B). The nature of vanishing rotations of gradient fields for simply
connected domains with boundary conditions on connected subsets of ∂B is employed as constraint to
make sure H is indeed a gradient field. Once H is known, the displacements can be independently
determined in the postprocessing step
ˆ
Π post = 1
2 ∇u : ∇u − ∇u : H dV ⇒ min . (2.2)
B u
Dirichlet boundaries are taken into consideration by the relation u = ū ⇔ H × n = ∇ū × n on ΓD ,

where n is the outward unit normal vector. To determine Π ext of (2.1) in terms of H the preprocessing
step ˆ ˆ
Π pre = 1
2 ∇g : ∇g − g · f dV − g · t dA ⇒ min (2.3)
B ΓN g
is solved analogously. In the 2D case (2.1) is equivalent to a Stokes-type problem [4] and the Lagrange
multiplier is sought in the standard solution space λ ∈ L2 (B). Suitable corresponding finite element
approximations for this are well known (see e.g. [3]) and in the following for H h and λh the Taylor
Hood type finite element discretizations P2H -P1λ and Q2H -Q1λ and the MINI element P 1BH − P 1λ
are considered. For solution spaces for the Lagrange multiplier and suitable discretizations in the 3D case
the reader is referred to [4]. For the discretization uh in (2.2) and g h in (2.3) H 1 -conforming Lagrange
finite element interpolation is used.
Penalty formulation: To further reduce the number of solution variables and consequently the number
of degrees of freedom a straightforward penalty approach can be applied. The integrand of the constraint
term associated with the Lagrange multiplier of (2.1) is then replaced by a penalty term 2h α
(rot H)2
with the scalar valued penalty parameter α and the mesh size parameter h. Consequently, H is the
only discretization variable in this approach. The corresponding finite elements are named P2H -pen for
simplical meshes and Q2H -pen for quadrilateral and hexahedral meshes.
3 Numerical examples
In this section the proposed finite element formulations are numerically evaluated with respect to conver-
gence behavior and numerical efficiency. For this, the gradient-enriched strain energy W (F , ∇F ) =
W loc (F ) + W nloc (∇F ) is chosen, in which W loc is a compressible Neo-Hooke energy function and
2
W nloc (∇F ) = µl2 ∇F 2 is a gradient enrichment term (c.f. [2, 11]) with the additional length param-
eter l. In what follows the elasticity constants E = 500 MPa, ν = 0.3 and l = 0.1 mm are chosen.
Unit square problem with smooth solution: The first numerical example is a unit square domain
B = (0, 1) × (0, 1) mm2 , for which the displacement solution field u = (uX , uY )T with uY = β(XY (X −
1)(Y − 1))2 and uX = 0 is chosen. For this, the essential boundary conditions u = 0 and ∇u = 0 on
ΓD = ∂B hold. Note, that β = −50 mm−8 is a scaling factor so that the order of magnitude of the
displacement field is within the nonlinear elastic range. The volume load f is computed analytically via
the strong form obtained by variation of (1.1) under plain strain assumption. Error measures are then
constructed from the finite element solution uh of (2.2) and H h of (2.1) respectively. As depicted in Figure
1 a) and b) convergence is observed for all elements both in the L2 -norm of the displacement error and in
28
0.100
10-2
0.010
0.001
10-4
∇u-Hh L2
u-u h L2
1.00
10-4 1.00
10-6 10-5
P1BH -P1λ P1BH -P1λ
2.02 10-6 2.03
P2H -P1λ P2H -P1λ
10-8 10-7
Q2H -Q1λ 2.00 Q2H -Q1λ 2.01
100 1000 104 105 100 1000 104 105

number of dof (main bvp) number of dof (main bvp)
a) b)
Figure 1. Unit square domain with smooth solution: a) Displacement error ||uh − u||L2 (B) and b) displacement
gradient error ||∇u − H h ||L2 (B) over numbers of degrees of freedom for uniform mesh refinement.
2.0
1.04
P2H -P1λ
relative displacement at point A
Q2H -Q1λ
1.02 1.5
Q2H -Q1λ *
Q2u -Q2H -Q1λ
1.00 ±0.5% 1.0 P2u -P2H -P1λ
0.98
0.5
0.96
0.05 0.10 0.50 1 0.0
a) average computing time t (s) b)
Figure 2. Comparative study on the unit square domain. Plot a) depicts the relative displacement uY /uY,ref with
uY,ref = −0.1953 mm at the point A= (0.5, 0.5) mm. Plot b) illustrates the corresponding computing
time for the converged mesh (marked with bullet points). Convergence criterion is ∆uY /uY,ref ≤ 0.1%.
the L2 -norm of the displacement gradient error. In order to evaluate computing efficiency, the proposed
P2H -P1λ and Q2H -Q1λ discretization is compared to the three-field formulations P2u -P2H -P1λ and
Q2u -Q2H -Q1λ (c.f. [7]). The displacement over computing time plot for uniform mesh refinement
(Figure 2) unveils a speed up factor of ≈ 2 of the proposed formulation over the three field formulations.
Taking into account the additional computing time of the pre- and postprocessing step, as denoted with
an asterisk for the Q2H -Q1∗λ formulation unveils a relatively small increase in computing time.
Cook’s problem in 3D In this numerical example, the penalty formulation is tested on the 3D Cook’s
problem (Figure 3 a). As the domain is clamped at X = 0 mm and loaded at X = 48 mm the Dirichlet
and Neumann boundary data is given with u|ΓD = 0 ⇔ (H ×n)|ΓD = 0 and t = (0, 0, p0 )T on ΓN . Figure
3 c) illustrates the influence of the micro-length parameter l on the stifness of the global system, since
the displacement response of the gradient elasticity formulations is lower than of the classical elasticity
elements P2u and Q2u .
4 Conclusion
A C 0 continuous finite element formulation for gradient elasticity at finite strains with a reduced number of
solution variables has been proposed. Numerical results in 2D show convergence and improved computing
efficiency of the proposed discretization compared to the three field mixed formulations (c.f. [7]). In the
3D case the penalty approach, which further reduces the number of solution variables whilst enabling
standard discretization, yields converging numerical results the ability to model size effects.
29
p0
16 24.5
24.0
displacement at point A
B
23.5
44
23.0
8 12 Q2H -pen
P2H -pen
22.5
Z Q2u local displacement formulation
P2u local displacement formulation
48 22.0
Y 500 1000 5000 1 × 104 5 × 104 1 × 105
X number of dof (main bvp)
(a) (c) (c)
Figure 3. Displacement convergence/ size effect study on the 3D Cook’s problem a) with p0 = 62.5 MPa. Plot b)
shows the contourplot of the uZ displacement field. Plot c) compares the displacement at the evaluation
point A=(48, 8, 60) mm over DOFs for classical displacement elements to the proposed gradient elasticity
elements with length parameter l = 3.5 mm. The penalty parameter is α = 103 .
Acknowledgments
The authors greatly appreciate financial funding by the German Science Foundation (Deutsche Forschungs-
gemeinschaft, DFG) in the Priority Program 1748 “Reliable simulation techniques in solid mechanics.
Development of non-standard discretization methods, mechanical and mathematical analysis”
References
[1] E.C. Aifantis. On the role of gradient in the localization of deformation and fracture. Int. J. Eng.
Sc. 30.10 (1992), pp. 1279–1299.
[2] H. Askes and E.C. Aifantis. Gradient elasticity in statics and dynamics: An overview of formulation,
length scale identification procedures, finite element implementations and new results. Int. J. Solids
Struct. 48 (2011), pp. 1962–1990.
[3] D. Boffi, F. Brezzi, and M. Fortin. Mixed Finite Element Methods and Applications. Springer, 2013.
[4] D. Gallistl. Stable splitting of polyharmonic operators by generalized stokes systems. Math. Comp.
86.308 (2017), pp. 2555–2577.
[5] R. Mindlin. Micro-structure in linear elasticity. Arch. Ration. Mech. Anal 16 (1964), pp. 51–78.
[6] M. Ortiz and G.R. Morris. C0 finite element discretizations of Kirchhoff’s equations of thin plate
bending. Int. J. Numer. Meth. Engng. (1988).
[7] J. Riesselmann, J. Ketteler, M. Schedensack, and D. Balzani. Three-field mixed finite element
formulations for gradient elasticity at finite strains. (accepted) (2019).
[8] S. Rudraraju, A. Van der Ven, and K. Garikipati. Three-dimensional isogeometric solutions to
general boundary value problems of Toupin’s gradient elasticity theory at finite strains. Comp.
Methods Appl. Mech. Engrg. 278 (2014), pp. 705–728.
[9] J.Y. Shu, W.E. King, and N.A. Fleck. Finite Elements For Materials With Strain Gradient Effects.
Int. J. Numer. Meth. Engng. 44 (1999), pp. 373–391.
[10] R. Toupin. Theories of elasticity with couple-stress. Arch. Ration. Mech. Anal 17 (1964), pp. 85–
112.
[11] N. Triantafyllidis and E.C. Aifantis. A gradient approach to localization of deformation I. Hypere-
lastic materials. Journal of Elasticity 16 (1986), pp. 225–237.
[12] L. Zybell, U. Mühlich, M. Kuna, and Z.L. Zhang. A three-dimensional finite element for gradient
elasticity based on a mixed-type formulation. Computational Materials Science 52 (2012), pp. 268–
273.
30

Determining the Constitutive Parameters of a Macro-

scale Second-Gradient Model for Planar Pantographic
Structures by Using Optimization Algorithms
1
Navid Shekarchizadeh ∗
and Masoud Abedi2
1 Department of Basic and Applied Sciences for Engineering, Sapienza University of Rome, Rome, Italy; International Research
Center for the Mathematics and Mechanics of Complex Systems-M&MOCS, University of L’Aquila, L’Aquila, Italy
2 Institute of Computer Science, University of Rostock, Rostock, Germany; Johann Heinrich von Thünen-Institut, Rostock, Ger-
many
Abstract. Pantographic structures have been proposed as a group of metamaterials that show high toughness
in extension. Modeling such structures is technically possible through microscopic first-order continuum
theories by using a suitably small length-scale, although the computational cost would be high. The aim of
this research is to estimate the constitutive parameters of a planar pantographic structure by means of an
optimization process. A previously-proposed macroscopic second-gradient model which is characterized by
deformation energy is used for modeling the mechanical behavior of the structure. The macroscopic model
will be developed based on the results of the numerical simulations of a microscopic model. In this problem,
an evolutionary multi-objective optimization algorithm is utilized which minimizes the squared error of the
outputs in order to determine the parameters of the panthographic structure.
Introduction
Metamaterials are referred to a group of materials that are designed carefully to have tailored properties.
Pantographic structures are an example of metamaterials that are proposed as a group of structures to
have high toughness and high strength-to-weight ratio which makes them attractive for industrial and
aerospace applications [4]. Figure 1 shows an example of a pantographic structure. The structure is made
up of straight beams which are connected together, at their intersection points, by cylinder pivots. These
pivots play a significant role in absorbing large amounts of energy by restricting the relative motion
of fibers. It is possible to model pantographic structures using classical continuum mechanics, which
is based on Cauchy theories, and describe the mechanical behavior in detail. However, it will not be
efficient in terms of computational cost. An appropriate solution is to employ “Generalized Continuum
Mechanics” in order to capture the effects of the substructure at a scale which is large enough to neglect
the geometrical complexities of the micro-structure.
The goal of the present work is to identify the constitutive parameters of a macro-scale model for a
planar pantographic structure in a way that the inspected properties of the structure matches well with
that of the data obtained, using Cauchy model, from an exact 3D model. This model will be a tool for
calculation and prediction of the mechanical behavior of the structure. As a continuation to perivous
studies on numerical identifications [1, 3, 4], here, an optimization algorithm is exploited for an automated
and easier parameter identification process.
Different forms of Genetic Algorithm (GA), which is among evolutionary algorithms, has been developed
for dealing with optimization problems, such as Non-Dominated Sorting Genetic Algorithm (NSGA) [7]
and NSGA-II [2]. The NSGA-II yields surprisingly better results compared the other two since it uses
two criteria for ranking solutions: grouping the objective function values in ’Fronts’ and calculating the
crowding distance for values of each ’Front’ [2].
In this research, we carry out the constitutive parameter identification of a planar pantographic structure
in two cases. First, the optimization is done with one ’fitness’ function using GA and then with two
objective functions using NSGA-II.
∗ Corresponding author: navid.shekarchizadeh@sbai.uniroma1.it ORCID iD: https://orcid.org/0000-0002-5750-7801
31
Figure 1. A pantographic structure and its geometrical parameters
1 Macro Model and Geometry
For describing the pantographic structure at macro level, the second-gradient model presented in [3] is
considered here. In this model, the current position x of a material point is a function of the position X
in the reference configuration through the map χ : Ω → R2 where Ω is a rectangular plane region.
x = χ(X) = (X α + uα (X β ))Dα (1)
The deformation gradient is F = ∇χ and the field of unit vectors tangent to the two families of fibers is
eα = FDα /||FDα || where Dα is a Lagrangean basis, α = {1, 2}. The auxiliary vector field cα is:
∇F|Dα ⊗ Dα
cα = where (∇F|Dα ⊗ Dα )β = ∂α Fαβ = ∂αα χβ (2)
||FDα ||
The strain measures, stretch of fibers (εα = F|Dα − 1), fiber curvature (κα = ||cα − (cα .eα )eα || =
||(I − eα ⊗ eα )cα ||), and shear distortion γ (sin(γ) = e1 .e2 ) are introduced. The strain energy density
that considers all the above contributions is assumed to be as follows.
X 1 1 1
WM (εα , κα , γ) = ( KI ε2α + KII κ2α ) + Kp γ 2 , KI , KII , Kp > 0, constant (3)
α
2 2 2
The governing equations are obtained using a variational principle:

Z
δ WM (εα , κα , γ)dΩ = 0 ∀ δu (4)
Ω
The pantographic structure shown in Fig. 1 is considered for modeling in this study. The geometrical
parameters, shown in Fig.1, are as follows: L = 0.2096 m, l = 0.06987 m, pb = 4.85 mm and the height
of each beam, hb = 0.9 mm, the width of each beam, bb = 1.6 mm, the height of each pivot, hp = 1 mm,
and the diameter of each pivot dp = 1.2 mm. The material used for the structure is Polyamide PA 2200
(Young Modulus = 1600 Mpa, Poisson’s ratio = 0.3, and mass density = 930 kg/m3 ). Bias extension tests
are simulated on the structure in COMSOL Multiphysics FEM enviroment and the total stored energy
and the angles ψ and ϕ are calculated. ψ is the angle between two beams intersecting at the center of
the specimen and ϕ is the bending angle of the beam that meets the corner of the specimen, see Fig. 1.
The reason of choosing these two angles is that ψ and ϕ are the representatives of the shear energy in
the central region and the bending energy in the most bent region respectively [3].
2 Optimization of the Macro Model Parameters
In this section, the considered results of the macro model are fitted with the micro model results through
an optimization process. The constitutive parameters of the macro model, KI , KII , and Kp , are deter-
mined by the optimization algorithm. At first, we consider the squared normalized error of the outputs
32
(The total stored energy and the two angles) as the ’fitness’ function (f1 ) and minimize it using GA. Then,

we defined another objective function, which grants double significance to ψ compared to ϕ Eq. (5) .
This is due to the fact that the energy involved in ψ angle is mainly governed by Kp [3] and on the other
hand by inspecting the energy components, we can see that the shear energy has the largest portion
of the total stored energy. Therefore, if the predicted ψ has lower error, the chance of predicting more
accurate energy values is higher. NSGA-II is utilized for the new optimization problem which has two
objectives. For investigating the effect of defining the new objective function, NSGA-II is modeled with
the same optimization algorithm parameters.
n
X n
X
ψM − ψm ϕM − ϕm
f2 = 2 ( i m i )2 + ( i m i )2 (5)
i
ψi i
ϕi
where i is the index of the i-th step loading and M and m denote the quantities from macro and micro
models, respectively.
In NSGA-II, the group of objective function value(s) that dominate other values are preferred, and they
are called the Pareto Front. The fronts are chosen using a matrix. The objective function values are
plotted on a graph with the axes showing the first and second objective function values. The advantage of
Pareto Front is that a set of optimum points can be assessed. The importance of having a set of optimum
solutions is that one can choose the optimum point even in case of occurring unexpected limitations [2].
Figure 2(a) shows the Pareto Front and the Fig.2 (b) shows the number of points in Pareto Front in the ten
first iterations. In complicated cases of multi-objective optimization problems, in case of applying NSGA-
II, the number of points in Pareto Front fall to zero after a few iterations and then suddenly it increases
as much as the number of all point. In this condition, NSGA-II acts poorly because of loss of selection
pressure in objective function evaluation. For overcoming this problem and increasing the accuracy
of ranking, the ’dominance’ concept was replaced by ’Strictly dominance’, ’Weakly dominance’, and
’Indifferent’ concepts through Fuzzy methods. Later, the Fuzzy-based Pareto Optimality was presented
using the Left Gaussian function and the Fuzzy-Pareto-Dominance [5, 6]. Since the plot of Pareto Front
is desired and the algorithm performs well in high iterations, our optimization has not encountered the
above-mentioned problem. The results show that the population is not high enough to have a Pareto
Front with less points than the population. On the other hand, considering the computational cost of
the model, it is not efficient to increase the population.
Figure 2. (a)The Pareto Front and (b)number of points in Pareto Front in every iteration
3 Results
By means of the optimization process, the constitutive parameters of the planar macro model are iden-
tified. In the first optimization, with one fitness function, the average value of the squared normalized
33
error of the optimum point was 1.455 and then it was slightly decreased by defining another objec-
tive function.The constitutive parameters calculated by the GA are KI = 6.274 × 105 (N/m), KII =
7.81 × 10− 2 (Nm), and Kp = 3.684 × 103 (N/m). In NSGA-II, since it is a multi-objective optimization,
a single set of parameters is not introduced as the optimum point, in other words, we have a number of
optimum points in the Pareto that non of them dominate the others, therefore, we may choose any of
the points with regard to the problem conditions. The values of the stored energy and the angles ψ and
ϕ calculated by simulating a bias extension test of the pantographic structure using the planar second-
gradient macro model are compared (in Fig. 3) with the values from Cauchy continuum micro model
which were previously calculated in [4] for the same structure. All the calculated values show acceptable
accordance except for the stored energy after 0.04m of displacement, where we see a discrepancy.
Figure 3. Comparison of the values of (a)Stored Energy, (b)The angle ψ, and (c)The angle ϕ from the micro[4]
and macro models
4 Conclusion
In this paper, a multi-objective optimization algorithm is utilized for identifying the constitutive param-
eters of a second-gradient macro model for a planar pantographic structure by minimizing the error of
the outputs of the model. By the virtue of the automated algorithm, the reduced-order model could be
created for different structures more easily.
References
[1] Michele De Angelo, Emilio Barchiesi, Ivan Giorgio, and B. Emek Abali. Numerical identification of
constitutive parameters in reduced-order bi-dimensional models for pantographic structures: appli-
cation to out-of-plane buckling. Archive of Applied Mechanics (2019), pp. 1–26.
[2] Kalyanmoy Deb, Samir Agrawal, Amrit Pratap, and Tanaka Meyarivan. A fast elitist non-dominated
sorting genetic algorithm for multi-objective optimization: NSGA-II. In: International conference
on parallel problem solving from nature. Springer. 2000, pp. 849–858.
[3] Francesco dell’Isola, Ivan Giorgio, Marek Pawlikowski, and Nicola Luigi Rizzi. Large deformations
of planar extensible beams and pantographic lattices: heuristic homogenization, experimental and
numerical examples of equilibrium. Proceedings of the Royal Society A: Mathematical, Physical and
Engineering Sciences 472.2185 (2016), p. 20150790.
[4] Ivan Giorgio. Numerical identification procedure between a micro-Cauchy model and a macro-
second gradient model for planar pantographic structures. Zeitschrift für angewandte Mathematik
und Physik 67.4 (2016), p. 95.
[5] Zhenan He, Gary G Yen, and Jun Zhang. Fuzzy-based Pareto optimality for many-objective evolu-
tionary algorithms. IEEE Transactions on Evolutionary Computation 18.2 (2013), pp. 269–285.
[6] Mario Köppen, Raul Vicente-Garcia, and Bertram Nickolay. Fuzzy-pareto-dominance and its appli-
cation in evolutionary multi-objective optimization. In: International Conference on Evolutionary
Multi-Criterion Optimization. Springer. 2005, pp. 399–412.
[7] Nidamarthi Srinivas and Kalyanmoy Deb. Muiltiobjective optimization using nondominated sorting
in genetic algorithms. Evolutionary computation 2.3 (1994), pp. 221–248.
34

Development of hybrid lattice materials using additive

manufacturing
Christina Völlmecke ∗1
, Vivian Dreßler1 , Daniel Melzer1 , Gregor Ganzosch2 and Anton Köllner1
1 Technische Universität Berlin, School V, Institute of Mechanics, Stability and Failure of Functionally Optimized Structures
Group, Einsteinufer 5, 10587 Berlin, Germany
2 Technische Universität Berlin, School V, Institute of Mechanics, Continuum Mechanics and Material Theory group, Einsteinufer
5, 10587 Berlin, Germany
Abstract. Additive manufacturing allows slender lattice materials to be developed very easily. These slender
structures are prone to buckling. By introducing pre-stressed fibres into the lattice, the buckling mode could
be changed and subsequently the critical load would increase. This has been shown theoretically for unit
cells (UC). Herein, the current strategies of the group for developing UCs using additive manufacturing are
presented.
Introduction
In large scale civil engineering applications, steel columns are common structural elements. Slender
steel columns that are additionally reinforced by pre-stressed cable stays generally exhibit an increased
strength in axial compression since buckling displacement is inhibited, e.g. [3, 4]. So far this concept has
only been applied on the large scale. Thus, the transferability of this concept to smaller length scales is
to be investigated. The potential development of novel material concepts had been shown by previous
analytical investigations [5]. Experimental investigations on the buckling and postbuckling behaviour
of simply-supported, pre-stressed stayed columns under pure axial compression have been conducted on
smaller length scales. They all displayed a reliable repeatability of the results, yielding a critical load
several times higher than the bare column whilst the postbuckling characteristics remain unchanged [1,
2]. Lately, the group has been working on the development of hybrid fibre lattice materials using additive
manufacturing. Therefore unit cells (UC) of the material were manufactured using two different printing
technologies, namely fused deposition modelling (FDM) and stereolithography (SLA), as will be presented
henceforth.
Figure 1. Schematic representation of a UC.
∗ Corresponding author: christina.voellmecke@tu-berlin.de
35
1 Manufacturing of hybrid lattice unit cells
Following up on the groups previous manufacturing approaches for pre-stressed hybrid lattice UCs (see
Fig. 1) two manufacturing strategies were chosen.
Strategy A implemented the fibre directly into the printing process whilst with strategy B two parts were
printed and adhesive boding of the two halves was undertaken. Both strategies were implemented using
FDM (Ultimaker 3 extended) as well as SLA (Formlabs Form 2) printing.
1.1 Fibre pull-out tests
In order to gauge the bonding of the fibre with the printing material prior to manufacture of UCs,
fibre pull-out tests were conducted using a microforce testing devices (MTS Tytron 250, MTS Systems
Corporation, Eden Prairie, USA). A special mounting system was designed to guarantee repeatability of
the tests (Fig. 2).
Figure 2. Fibre pull-out test.
Different filament materials were used with the FDM printer such as Polylactic Acid filament (PLA),
Polymethylmethacrylat (PMMA) and Acrylnitril-Butadien-Styrol (ABS). For the SLA printed formlabs
clear resin was used. As fibre materials glass, carbon and basalt were taken. Adhesive bonding was
undertaken using acetone, modelling glue, super glue and epoxy resin. From the fibre pull-out tests it
can at this stage be deducted that strategy A (direct implementation of the fibre) results in insufficient
bonding of the fibre with the material. Using adhesive bonding of the two halves generally resulted in
better bonding of the fibre but is still to be developed further in order to achieve reliable results.
1.2 Mounting device for the UC
A special mounting device was developed for the manufacturing of the UCs (Fig. 3. The device is designed
such that pre-stress can be applied. Using this device a UC was successfully manufactured as can be seen
in Fig. 4.
2 Conclusion
A novel approach on developing UCs for hybrid lattice materials using additive manufacture has been
presented. Two strategies were pursued, namely direct integration of the fibre in the printing process or
subsequent adhesive boding of two halves of the UC whilst implementing the fibre. Fibre-pull out tests
revealed that both strategies cannot yet guarantee perfect bonding of the material with the fibre. Hence
stresses will not to be transmitted sufficiently via the fibres. Thus, in order to implement the concept
of pre-stressed cable stays already used in large scale applications on a material level, the development
strategies need to be reviewed and amended.
36
Figure 3. Mounting device for UC.
Figure 4. UC made with formlabs clear resin and basalt fibre using strategy A.
References
[1] G. Ganzosch, M. Todt, A. Köllner, and C. Völlmecke. Experimental investigations on pre-stressed
stayed columns on smaller length scales. In: Proceedings of the Eighth International Conference on
Thin-walled Structures. 2018, pp. 1–15.
[2] A. Köllner, M. Todt, G. Ganzosch, and C. Völlmecke. Experimental and numerical investigation on
pre-stressed lattice structures. Thin-Walled Structures (under revision).
[3] D. Saito and M.A. Wadee. Post-buckling behaviour of prestressed steel stayed columns. Engineering
Structures 30.5 (2008), pp. 1224 –1239.
[4] M.A. Wadee, L. Gardner, and T.A. Hunt. Buckling mode interaction in prestressed stayed columns.
Proceedings of the Institution of Civil Engineers - Structures and Buildings 166.8 (2013), pp. 403–
412.
[5] C. Zschernack, M. A. Wadee, and C. Völlmecke. Nonlinear buckling of fibre-reinforced unit cells of
lattice materials. Compos Struct 136 (2016), pp. 217–228.
37
38
MS 02: Peridynamics and other non–Local Methods for Computational Mechanics

A nite strain microplane-plasticity model for concrete

Bobby Rio Indriyantho1 and Michael Kaliske ∗ 1
1 Institute for Structural Analysis, Technische Universität Dresden, 01062 Dresden, Germany
Abstract. A plasticity model is one of the powerful approaches which is essential for achieving softening
response of materials, besides damage mechanics theories. Meanwhile, the microplane approach has been
studied extensively and its development proved to be appropriate for modelling quasi-brittle materials, such as
concrete. Numerous constitutive laws of the microplane-plasticity model have been developed in many versions
for the small strain framework, and they have been utilised successfully to predict the load-displacement as
well as stress-strain curves including the post-peak behaviour. However, at high hydrostatic pressure, a plastic
response with very large strains or deformations may occur in concrete materials or building structures. Due to
this phenomenon, the extension of the microplane-plasticity model from the small strain to the nite strain case
is inevitable and it is required in order to represent the largely deformed behaviour of structures. The elastic-
plastic microplane approach based on the volumetric-deviatoric decomposition employing the Drucker-Prager
yield criterion extended to the nite strain regime is implemented in this present contribution. Moreover,
for eliminating the pathological mesh sensitivity and obtaining a more stable nite element solution due
to strain localization, the so-called implicit gradient enhancement, which is straightforward to be applied
in general nite element codes, is used here. Furthermore, numerical example in quasi-static case is given
and discussed to evaluate the proposed formulation. Finally, the simulated results are compared to existing
experimental investigations in order to evaluate and validate the accuracy of the newly proposed model.
Introduction
A plasticity model is one of the essential approaches to solve issues regarding strain softening of materials,
besides damage mechanics theories. Meanwhile, the so-called microplane model has been proven to be a
powerful constitutive approach for modelling quasi-brittle materials, such as concrete, since it was rstly
introduced by [2, 3]. The other microplane formulation is proposed by [7] based on the volumetric-
deviatoric (V-D) split and the kinematic constraints. However, strain localization arises and the nite
element solution yields unstable. In order to overcome these challenges, [8] developed an implicit gradient
enhancement employing nonlocal elds which is straightforward to be implemented in general nite
element codes.
Numerous versions of the microplane model for concrete have been developed and improved successfully
at small strains. Nevertheless, in certain loading cases, small strain microplane models produce inaccurate
results in terms of load-displacement or stress-strain curve as well as an incorrect damage visualization.
At high hydrostatic pressure, the response of concrete becomes plastic and the material exhibits very
large deformations with no observable cracks or damage. This experimental investigation in [4] has been
conrmed in [6] and concludes that the microplane damage formulation at the small strain framework
cannot yield good results for the aforementioned case. Consequently, since the microplane approach at
small strains is not sucient to accommodate very large deformations of concrete and the plastic response
occurs, the use of the microplane-plasticity model at nite strains is inevitable.
The present contribution is focused on the microlane model for concrete accommodating the plastic
response at very large deformations. The gradient-enhanced microplane-plasticity approach at small
strains in [9] is extended to the nite strain model using the Green-Lagrange strain tensors. The aim of
this work is to introduce and formulate the new extended gradient-enhanced microplane-plasticity model
within the nite strain framework for concrete to obtain a reliable and accurate response in quasi-static
case. Finally, the results are compared to the existing experimental data in order to evaluate the accuracy
of the newly proposed model.
∗ Corresponding author: michael.kaliske@tu-dresden.de
39
1 Finite strain microplane-plasticity formulation
An additive volumetric-deviatoric split using the Green-Lagrange strain tensor E has been employed in
[1, 5, 6] to extend the small strain model to the nite strain framework for concrete materials, so that it
enables the elastoplastic formulation at nite strains being easy to handle, since its equations are similar
to the small strain model as follows
EVe = EV − EVp , (1)
E eD = E D − E pD . (2)
Hence, the microplane free-energy function with hardening H and its variable κmic is
1 mic 1 2
Ψmic (E e ) = K (EV − EVp )2 + Gmic (E D − E pD ) · (E D − E pD ) + Hκmic . (3)
2 2
According to [5], however, the resulting stress should be the so-called co-rotated Cauchy stress tensor s
instead of the second Piola-Kirchho stress tensor S, since the microplane orientation should rotate with
the material, see [5] for details. By deriving the microplane free-energy Ψmic in the V-D split with respect
to the Green-Lagrange strain tensor E and, then, integrating it over all microplanes, the homogenised
co-rotated Cauchy stress tensor is
Z Z
3 3
s= K mic (EV − EVp ) U V U dΩ + 2Gmic U Dev T · (E D − E pD ) U dΩ, (4)
4πJ Ω 4πJ Ω
with
1 2 −1
V = J3C , (5)
3
1
Dev = n · Idev = n · Isym − n · C −1 ⊗ C, (6)
3
where V and Dev are the volumetric and deviatoric projection tensors, respectively. Finally, the true
(Cauchy) stress σ is obtained by
σ = RsRT , (7)
where R is the rotation tensor obtained by the polar decomposition of deformation gradient F.
2 Implicit gradient enhancement and yield function
The implicit gradient enhanced method is applied by modifying the Helmholtz equation for dening
nonlocal elds as follows
κ̄m − c∇2 κ̄m = κm , (8)
with

κm = max21
mic=1 κ
mic
, (9)
where the gradient parameter c controls the interaction range of the nonlocal part and ∇2 is the Laplace
operator. Moreover, κm and κ̄m are the local variable and its nonlocal counterpart, respectively. Using Eq.
(9), only one of 21 microplanes is utilised to perform the regularisation as adopted from [9]. Consequently,
its largest quantity is taken and updated in each time step.
By introducing the eect of the nonlocal eld κ̄m within the damage variable ω(κ̄m ) for obtaining the
softening response as implemented in [9], the modied Drucker-Prager yield function is
r
3 mic
F mic = S D · S mic mic
D + αSV − (1 − ω(κ̄m )) S0 + Hκmic ≤ 0, (10)
2
with
ω(κ̄m ) = αp (1 − exp(−βκ̄m )) , (11)
where αp and β govern the maximum degradation of the yield stress and the softening branch, respectively.
40
3 Numerical example
A tube squash test, based on the experimental investigation in [4], can show very large strains with plastic
responses in concrete at high hydrostatic pressure. Therefore, this test will be simulated using the newly
proposed model and its result will be compared each other. A hollow steel tube with an inner diameter
of 47.8 mm (76.2 mm outer diameter) and a height of 88.9 mm is lled by concrete and tested at the
axial constant displacement rate of 0.0254 mm/s. All material parameters used here are shown in Table
1. The steel tube is modelled by common von Mises plasticity and assumed as a perfect bond between
steel and concrete surface. The calibration of the empty steel tube is needed before performing the tube-
squash test, see [6] for details. Afterwards, the steel-concrete tube is simulated but only one-forth of the
specimen is discretised, due to the symmetry, using 4305 nodes and 2560 elements.
Table 1. Material parameter of a tube-squash test.

Parameter Filled concrete Steel tube
E [MPa] 20000 200000
ν [-] 0.2 0.3
S0 [MPa] 30.0 -
H [MPa] 800 500
α [-] 0.5 -
β [-] 500 -
αp [-] 0.998 -
2
c [mm ] 500 -
σy [MPa] - 360
As can be seen in Figure 1, the load-displacement curve simulated by the proposed model has a good
agreement compared to the experimental result in [4]. The nal axial displacement obtained by this
squash test is about 31 mm or about 1/3 of the tube height. It indicates that very large strain occurs in
this simulation, since the axial strain amounts approximately up to 34%.
Figure 1. Comparison of load-displacement curve for experiment and simulation.
In the necessity of plastic zone visualisation, 21 microplane hardening variables κmic are homogenised
hom
and a scalar hardening measure κ is taken into account as
3
R
hom 4π κmic dΩ
ΩR
κ = 3 . (12)
4π Ω
dΩ
From Figure 2(a), one can see the evolution of plastic response at dierent axial displacement uz up to
nal time step. In this tube-squash test simulation, the lled concrete becomes plastic in the middle
part of concrete. It also conrms the experimental investigations in [4] that no damage can be observed
due to plastic behaviour, see Figure 2(b). By the load-displacement curve and plastic visualisation, the
proposed model can predict the responses of concrete in the tube-squash test accurately.
41
(a) (b)
uz = 15.21 mm uz = 22.83 mm uz = 31.33 mm

0 0.1 0.2 0.3 0.4 0.55
Figure 2. Simulation results: (a) evolution of plastic zone, (b) comparison between simulation and experiment.
4 Conclusions
The nite strain microplane-plasticity approach yields good results for modelling concrete at high hydro-
static pressure in a tube-squash test. The simulated result conrms the tested specimen experimentally
and it can predict an accurate plastic response. Furthermore, more examples in static and dynamic cases
will be performed at future work.
Acknowledgments
The authors gratefully acknowledge the nancial support of Sächsiches Landesstipendium (Saxon Schol-
arship Program) with reference number L-201653.
References
[1] Z.P. Baºant. Finite strain generalization of small-strain constitutive relations for any nite strain
tensor and additive volumetic-deviatoric split. International Journal of Solids and Structures 33
(1996), pp. 28872897.
[2] Z.P. Baºant and P.G. Gambarova. Crack shear in concrete: crack band microplane model. Journal
of Structural Engineering 110 (1984), pp. 20152035.
[3] Z.P. Baºant and B.H. Oh. Microplane model for concrete for fracture analysis of concrete structures.
In: Proceedings of Symposium on the Interaction of Non-nuclear Munitions with Structures. U.S.
Air Force Academy. Colorado Springs, 1983, pp. 4953.
[4] Z.P. Baºant, J.J.H. Kim, and M. Brocca. Finite strain tube-squash test of concrete at high pressures
and shear angles up to 70 degrees. ACI Materials Journal 96 (1999), pp. 580592.
[5] Z.P. Baºant, M.D. Adley, I. Carol, M. Jirásek, S.A. Akers, B. Rohani, J.D. Cargile, and F.C. Caner.
Large-strain generalization of microplane model for concrete and application. Journal of Engineering
Mechanics 126 (2000), pp. 971980.
[6] B.R. Indriyantho, I. Zreid, and M. Kaliske. Finite strain extension of a gradient enhanced microplane
damage model for concrete at static and dynamic loading. Engineering Fracture Mechanics (2019).
doi: https://doi.org/10.1016/j.engfracmech.2019.106501.
[7] M. Leukart. Kombinierte anisotrope Schädigung und Plastizität bei kohäsiven Reibungsmaterialien.
PhD thesis. Institut für Baustatik, Universität Stuttgart, 2005.
[8] R.H.J. Peerlings, R. de Borst, W.A.M. Brekelmans, and J.H.P. de Vree. Gradient enhanced damage
for quasi-brittle materials. International Journal for Numerical Methods in Engineering 39 (1996),
pp. 33913403.
[9] I. Zreid and M. Kaliske. An implicit gradient formulation for microplane Drucker-Prager plasticity.
International Journal of Plasticity 83 (2016), pp. 252272.
42

A nonlocal ductile damage model for high strength

steels under dynamic multiaxial loading conditions
Marvin Nahrmann1 and Anton Matzenmiller ∗ 1
1 Institute of Mechanics, Dept. of Mechanical Engineering, University of Kassel, Mönchebergstr. 7, 34125 Kassel, Germany
Abstract. The high strength steel alloy HX340LAD (ZStE340) is investigated under quasistatic and dynamic
loading conditions for different flat specimens. A thermo-viscoplastic constitutive model is applied to capture
the temperature and strain rate dependent material characteristics. In order to consider the post-critical
material behaviour, the model is enhanced with a nonlocal ductile damage approach. The predicted results
exhibit a close agreement to the measured values leading to a successful verification of the model.
1 Introduction
Today’s multi-material designs of lightweight car body structures are built upon a broad variety of
different steel and aluminium alloys. The diverse material properties of these metals are tailor-made for
the respective application areas. In order to ensure the occupant safety, a high tensile strength for the
protection of passengers is required as well as a high ductility absorbing the crash energy in the crumpling
zone by means of large plastic deformations. Microalloyed high strength steels (HSLA steels) provide a
good material strength and ductility and are well suited for complex cold-formed vehicle body parts. Due
to dynamic multiaxial loading conditions in the case of metal forming or vehicle crash, the finite element
simulation of HSLA steel is still a sophisticated task. Particularly for high strain rates, the conversion of
dissipated work into heat causes a locally significant temperature rise. Hence, a temperature and strain
rate dependent constitutive model is applied for a thermo-mechanical coupled simulation. Moreover,
a nonlocal ductile damage model is appended to capture the mesh independent post-critical material
behaviour and enable a good prediction of the entire energy absorption potential.
2 Constitutive model for thermo-viscoplastic material behaviour
The temperature and strain rate dependent material behaviour of high strength steels is considered by
the thermo-viscoplastic constitutive model comprehensively presented in [10]. The basic framework of the
model is adopted from [4] and [5], whereby it is originally applied to the simultaneous hot/cold forging of
metals–see [7]. In this contribution, only a moderate temperature range, from room temperature up to
approximately 600 K is considered, which is sufficient to comprise the adiabatic heating for high strain
rates. The constitutive equations of the thermo-viscoplasticity model are implemented as a user defined
material model (*MAT_USER_DEFINED_MATERIAL_MODELS) into the commercial FE software
LS-DYNA and include the temperature dependency of the Young’s modulus, the initial yield stress, and
the nonlinear isotropic hardening.
Adiabatic heating is taken into account by the Taylor-Quinney approximation, which determines the
temperature increase by means of the viscoplastic stress power and the specific heat capacity c.
γTQ
θ̇ = T · Ėvp , γTQ = 0.90 (1)
ρc
Here, the coefficient γTQ is specified to convert 90 % of the viscoplastic work into heat. The resulting
thermo-mechanical coupled problem is solved by the staggered solution scheme, whereby the thermal
problem is solved with a constant displacement field and the mechanical problem with fixed temperatures,
respectively.
∗ Corresponding author: post-structure@uni-kassel.de
43
3 Nonlocal ductile damage approach
The thermo-viscoplastic material model is coupled with a ductile damage approach based on continuum
damage mechanics allowing the prediction of damage and failure caused by large plastic deformations.
The concept of effective stresses is employed in order to obtain the Cauchy stress tensor in the damaged
(physical) space a T on the basis of the damage variable D and the Cauchy stress tensor in the undamaged
(effective) space T. All internal strain variables shall be identical in both configurations, due to the strain
equivalence principle–see [5]. The ductile damage growth is driven by the equivalent viscoplastic strain
Ēvp , whereby the approach of Lemaitre in [6] is applied, introducing the model parameter εc related to
the damage initiation and εf for failure. The exponent nD controls the course of the damage evolution.
n
a Ēvp − εc D
T = (1 − D) T , D = , 0≤D<1 (2)
εf − εc
The nonlocal option *MAT_NONLOCAL is inbuilt in LS-DYNA and based on the integral approach
presented in [11] preventing the localisation of an arbitrary spatial field. In this contribution, the nonlocal
formulation is applied to the damage variable D to maintain mesh independent outcomes–see also the
˙
investigations in [2]. Hence, the rate of the nonlocal damage variable D̃(x) at the reference material point
x is obtained by weighting the local damage rate Ḋ(y) at neighbouring points y over a certain region Ω
of the continuum.
Z Z
˙ 1
D̃(x) = Ḋ(y) w(x, y) dy , W (x) = w(x, y) dy (3)
W (x) Ω Ω
For the implementation in conjunction with finite elements, the integral formulation is approximated
by sums over the elements included in the integration region Ω that is determined by the characteristic
length lc –see Fig. 1 (left).
"N #" N #−1 p −q
˙ X X kx − yi k
D̃(x) ≈ Ḋi wi Vi wi Vi , wi = w(x, yi ) = 1 + (4)
i=1 i=1
lc
Basically, a bell-curved function is assumed for weighting the local variable. The weighting function
w(x, yi ) is evaluated at yi for the elements i = 1, ..., N in the integration region and is hard-coded in
LS-DYNA with two model parameters, p and q, specifying its shape. Here, the values q = 2 and p = 8
are selected, leading to a broad localisation zone within the integration region. Note that this function
is only valid in the integration area, so if kx − yi k/lc ≤ 1–see Fig. 1 (right).
In order to enable the nonlocal formulation, the characteristic length lc has to be greater than the
element length in the damage zone. However, the determination of the characteristic length is still a
sophisticated challenge. A measurement method for concrete is outlined in [3] on the basis of notched
and unnotched tensile tests. In contrast to concrete, the material behaviour of high strength steels is
much more ductile. Furthermore, it is well known that the damage and failure behaviour highly depends
on the stress state–see [12]. Discussions on the internal length for ductile metals are made in [2] and [1],
nevertheless, a general measurement technique has not established yet at all. In [13], the characteristic
length is expected to relate to the damage mechanism. In the case of ductile metals, two different damage
and failure mechanisms occur. The first is dimple rupture under hydrostatic loading, whereby internal
voids are expanded and void coalescence leads to failure. The second damage mechanism is shear failure
by means of void elongation within a narrow shear band. Thus, the nonlocal damage model is supposed
to consider stress state dependency for the simulation of structures under multiaxial loading conditions.
q = 2, p = 8
w(x, yi )
kx − yi k/lc
Figure 1. Integration region Ω determined by the characteristic length lc (left), weighting function w (right)
44
4 Verification results for a high strength steel under dynamic multiaxial loading conditions
The thermo-viscoplastic material model with nonlocal ductile damage is verified for six flat specimens,
made of the microalloyed high strength steel HX340LAD (ZStE340), for quasistatic (qs) as well as dynamic
loading conditions under low (25 mm/s), medium (250 mm/s) and high velocities (2500 mm/s)–see
Fig. 2. All model parameters of the thermo-viscoplasticity model are identified and listed in [10], whereby
tensile test data from [9] and [8] as well as shear test data from [12] are employed. The parameters for
the ductile damage approach and the nonlocal formulation are presented in Fig. 2 for all investigated
specimens. Different damage model parameters are permitted for each specimen in the first instance,
due to the absence of a stress state dependent nonlocal damage model. The characteristic length for
every specimen is defined on the basis of the plastic zone prior to failure, leading to a similar extent of
lc = 1
εc = 1.25
εf = 1.46
lc = 3
εc = 0.50
εf = 0.60
lc = 1
εc = 1.51
εf = 1.72
lc = 4
εc = 0.80
εf = 0.94
lc = 2 lc = 2.5
εc = 0.40 εc = 0.95
εf = 0.55 εf = 1.14
Figure 2. Simulation results compared to test data ([9],[12]) for the perforated, unnotched and notched tensile
specimen (left column) as well as for the -20◦ , 0◦ and 45◦ shear specimen (right column)
45
Figure 3. Spatial fields of the equivalent plastic strain and the damage variable prior to failure for the -20◦ , 0◦
and 45◦ shear specimen as well as for the notched, perforated and unnotched tensile specimen
high plastic deformations and damaged material–see Fig. 3. This results in noticeable different values for
lc , comparing tensile and shear test for example. The overall simulation results are in good correlation
with the according test data for nearly all specimens and loading conditions. An interesting effect is the
reduction of elongation for shear dominated high speed loading, recognisable in the experimental as well
as in the predicted stress-strain courses–see right centred diagram in Fig. 2. A probable reason is the
strong adiabatic heating in a narrow band resulting in thermal softening and premature failure.
References
[1] F. R. Ahad, K. Enakoutsa, K. N. Solanki, and D. J. Bammann. Nonlocal modeling in high-velocity
impact failure of 6061-T6 aluminum. International Journal of Plasticity 55 (2014), pp. 108–132.
[2] F. X. C. Andrade, J. M. A. Cesar de Sa, and F. M. Andrade Pires. Assessment and comparison of
non-local integral models for ductile damage. Int J of Damage Mech 23.2 (2014), pp. 261–296.
[3] Z. Bazant and G. Pijaudier-Cabot. Measurement of Characteristic Length of Nonlocal Continuum.
Journal of Engineering Mechanics 115.4 (1989), pp. 755–767.
[4] C. Bröcker and A. Matzenmiller. An enhanced concept of rheological models to represent nonlinear
thermoviscoplasticity and its energy storage behaviour. Cont Mech Therm 25 (2013), pp. 749–778.
[5] C. Bröcker and A. Matzenmiller. On the generalization of uniaxial thermoviscoplasticity with dam-
age to finite deformations based on enhanced rheological models. Technische Mechanik 34 (2014),
pp. 142–165.
[6] J. Lemaitre. A Continuous Damage Mechanics Model for Ductile Fracture. Journal of Engineering
Materials and Technology 107 (1985), pp. 83–89.
[7] A. Matzenmiller and C. Bröcker. Thermo-mechanically coupled FE analysis and sensitivity study
of simultaneous hot/cold forging process with local inductive heating and cooling. International
Journal of Material Forming 5 (2012), pp. 275–300.
[8] G. Meschut and M. Gerkens. Entwicklung einer Methode zum Nachweis der Einsetzbarkeit des
Hochgeschwindigkeits-Bolzensetzen unter Berücksichtigung der Bauteileigenschaften. Europäische
Forschungsgesellschaft für Blechverarbeitung e.V., report no. 455. 2017.
[9] G. Meschut, O. Hahn, D. Hein, A. Matzenmiller, and A. Nelson. Experimentelle und numerische
Untersuchungen des Crashverhaltens hybridgefügter Verbindungen. FOSTA report P958. 2015.
[10] M. Nahrmann and A. Matzenmiller. Modelling and identification of coupled thermo-viscoplastic
material behaviour with nonlocal ductile damage. Proceedings of the 12. European LS-DYNA Con-
ference, Koblenz. 2019.
[11] G. Pijaudier-Cabot and Z. Bazant. Nonlocal Damage Theory. Journal of Engineering Mechanics
113.10 (1987), pp. 1512–1533.
[12] A. Trondl, S. Klitschke, W. Böhme, and D.-Z. Sun. Verformungs- und Versagensverhalten von
Stählen für den Automobilbau unter crashartiger mehrachsiger Belastung. FAT-Schriftenreihe 283.
Forschungsvereinigung Automobiltechnik e.V. (FAT). 2015.
[13] V. Tvergaard and A. Needleman. Nonlocal effects on localization in a void-sheet. International
Journal of Solids and Structures 18 (1997), pp. 2221–2238.
46

Analysis of microwire torsion using a nonlocal

dislocation based plasticity formulation
1
Kolja Zoller ∗
and Katrin Schulz1
1 Institute
for Applied Materials - Computational Materials Science (IAM-CMS), Karlsruhe Institute of Technology, Kaiserstr.12,
76131 Karlsruhe, Germany
Abstract. The microstructural evolution of face-centered cubic microwires is studied using a physical mo-
tivated, homogenized continuum model of crystal plasticity. The dislocation configuration in the three-
dimensional space is thereby described via a Continuum Dislocation Dynamics (CDD) theory including a
dislocation source term. The resulting spatial distribution of dislocation densities and strain components are
shown for a relaxation problem with torsional loading.
Introduction
Microwires under torsion show a pronounced size effect [1], whereby the consideration of single crystal
allows to study the influence of strain gradients independent from grain sizes. Describing dislocation
densitiy as a homogenized ensemble of dislocation lines, the CDD theory [2] is suitable for a physical
consideration of such problem providing the benefits of a continuum formulation. A numerical algorithm
applied to pile-ups at grain boundaries is introduced in [6]. Additional, an extension including a dislo-
cation density production term for bending is formulated in [4] and for torsion in [7]. The connection
between pile-ups and size effects is discussed in [5].
1 Dislocation density based continuum model
The considered model describes the elasto-plastic deformation behavior of face-centered cubic single
crystal metals, whereby the plasticity solely results from the evolution of the dislocation microstructure
characterized by CDD densities. Operating on the mesoscopic scale, the model can be divided into two
coupled problems using the same length scale: The elastic problem calculating the stress field σ for a
given plastic state and the dislocation problem describing the dislocation evolution for a given stress field
resulting in plasticity via the plastic slip γs on all twelve slip systems S = 12. Each slip systems s is
characterized by its plane normal ms and the Burger’s vector bs = bs ds with the length bs and the slip
direction ds . The governing equations of both problems can be found in Tab. 1. Regarding the boundaries,
it has to be remarked, that the dislocations can leave the continuum at the surfaces unhindered. The
considered degrees of freedom are three nodal displacements plus four nodal dislocation density based
quantities per slip system.
The curvature production term q̇sprod enables the model to increase the number of dislocation loops in
the system. The production term is activated if the local stress exceeds a critical value and accounts
for a relaxation of the local stress to its critical state. The macroscopic ideal plastic material behavior
functions thereby as a upper limit for the global production rate.
47
Table 1. Set of governing equations in the three-dimensional space

Elastic problem Dislocation problem
Decomposed distorsion tensor Du = β el + β pl Orowan equation γ̇s = vs bs ρs n o
P
Plastic distorsion β pl = S s=1 γ s M s Velocity law vs = bBs sgn(τseff ) max 0, |τseff | − τsy
Schmid tensor M s = ds ⊗ ms Effective stress τseff = σ · M s − τsb
Infinitesimal total strain ε = sym Du Backstress τsb = Dµb s
∇ · κ⊥
s , κ⊥s = κs × ms
ρs pP
Stress-strain relation σ = C[ε − εpl ] Yield stress y
τs (ρ) = aµbs ρ
n n
Macroscopic balance equation − div σ = f B Dislocation density ρ̇s = −∇ · vs κ⊥ s + vs qs
Dirichlet boundary conditions u = uD GND density κ̇s = ∇ × (ρsvs ms )
Neumann boundary conditions σn = tN Curvature density q̇s = −vs ∇ · ρqss κ⊥ s + q̇sprod

κs κs
κ⊥ κ⊥
− 12 ρs + |κs | |κ s|
⊗ |κs |
+ ρ s − |κ s | |κ⊥
s
⊗ |κ⊥
s
· ∇ 2 vs
s | s |

Boundary condition vs |n × ms |ρs + (n × ms ) · κs = 0
The used system configuration is shown in Fig. 1. A Table 2. Material parameter

microwire with an aspect ratio of one to two is exam- µ = 27 GPa
ined. The bottom is fixed and the top is subjected ν = 0.347
to a twist deformation. The surfaces on the side are b = 0.286 µm
traction-free. We consider an instantaneous loading B = 10−4 ns GPa
with an subsequent dislocation relaxation. The re- D = 0.255 µm−3
sulting stress field consists mainly of two shear com- a = 0.35 µm−2
ponents with a linear distribution. Therefore, the
center is stress-free and the highest stress occurs at
the surface. The chosen material parameter in Tab. 2
are based on aluminium with a [0 1 0] crystal orien-
tation. Elastic isotropy is assumed.
Figure 1. Geometry
2 Relaxation of micro-torsion
The initial condition is dislocation free and the stresses at the surface are initially twice the critical stress
for the source activation. For the elastic problem a standard FEM solver with linear ansatz functions is
used. The flux-based dislocation transport equations considering dislocation mean-field stresses are solved
by a Runge-Kutta discontinuous Galerkin scheme with upwind flux, an implicit midpoint rule for the
evolution equations and constant ansatz functions. Both solvers operate on the same mesh consisting of
about 170 000 tetragonal elements. The spatial distribution over the radius of the resulting microstructure
is characterized by Fig. 2 and Fig. 3.
Figure 2. Resulting spatial distribution of the to- Figure 3. Resulting spatial distribution of the total
tal dislocation density ρ as well as screw strain ε and its elastic part εel and plastic
part κ1 and edge part κ2 of the GND part √
εpl over the normalized radius with
density
√ over the normalized radius with r = x2 + z 2 and R = 12 µm measured
r = x2 + z 2 and R = 12 µm measured in the middle of the microwire.
in the middle of the microwire.
∗ Corresponding author: kolja.zoller@kit.edu
48
The expanding dislocation loops produced near the surface in the initial step are leaving the continuum at
the surfaces and moving accordingly to the stress field towards the center forming a pile-up. Accordingly,
the piled-up total dislocation density consists mainly of geometrically necessary dislocation (GND) den-
sity. Due to the orientation of the slip systems and loading case, the screw part plays with a percentage
of 78.8 % of the GNDs a dominant role.
The total strain has a linear profile over the radius. The resulting plastic strain is nearly constant near
the surface and decreases from a breaking point linear towards the center in the same manner as the
total strain. The position of the break point correlates with the change in the dislocation density. The
elastic strain and therefore the shear stress is almost zero near the center and increase linear from the
break point towards the surface with the same gradient as the total strain.
3 Discussion
With the critical thickness theory (CTT) there are analytical studies on this topic [3]. When the strain
reaches a critical state, the area under the elastic strain curve remains constant in the following. The
qualitative distribution found by CTT is identical to the simulation results in this contribution. For the
given load even the quantitative distribution is similar except for a constant higher dislocation production
by a factor of one third. This is due to the three-dimensional consideration with an increasing influence
over the radius since each element has the same weight but the considered area increase with the radius.
One important point is that we wouldn’t expect such a torsion behavior showing the plastic plateau for
a classical continuum approach. However, the flow terms of the CDD formulation enable the motion of
the dislocation density produced near the surface into the center and therefore the plastification of the
material with a local stress smaller than the critical value.
Acknowledgments
The financial support for this work in the context of the DFG research group FOR1650 ‘Dislocation based
Plasticity’ as well as the support by the European Social Fund and the state Baden-Wuerttemberg is
gratefully acknowledged. This work was performed on the computational resource ForHLR II funded by
the Ministry of Science, Research and the Arts Baden-Württemberg and DFG.
References
[1] N. A. Fleck, G. M. Muller, M. F. Ashby, and J. W. Hutchinson. Strain Gradient Plasticity: Theory
and Experiment. Acta Metallurgica et Materialia 42(2) (1994), pp. 475–487.
[2] T. Hochrainer, S. Sandfeld, M. Zaiser, and P. Gumbsch. Continuum dislocation dynamics: towards
a physical theory of crystal plasticity. Journal of the Mechanics and Physics of Solids 63 (2014),
pp. 167–178.
[3] D. Liu, X. Zhang, Y. Li, and D. J. Dunstan. Critical thickness phenomenon in single-crystalline wires
under torsion. Acta Materialia 150 (2018), pp. 213–223.
[4] S. Schmitt, M. Stricker, P. Gumbsch, and K. Schulz. A mechanism-based homogenization of a dislo-
cation source model for bending. Acta Materialia 164 (2019), pp. 663–672.
[5] K. Schulz, M. Sudmanns, and P. Gumbsch. Dislocation-density based description of the deformation
of a composite material. Modelling and Simulation in Materials Science and Engineering 25(6)
(2017), p. 064003.
[6] K. Schulz, L. Wagner, and C. Wieners. A mesoscale continuum approach of dislocation dynamics
and the approximation by a Runge-Kutta discontinuous Galerkin method. International Journal of
Plasticity 120 (2019), pp. 248–261.
[7] K. Zoller and K. Schulz. A bottom-up continuum approach of crystal plasticity for the analysis of
fcc microwires under torsion. Proceedings in Applied Mathematics and Mechanics PAMM28 (2019),
10.1002/pamm.201900032.
49
50
MS 03: Current Trends in Computational Fracture Mechanics

A phase-field model for mixed-mode fracture

1,2
Meng Fan ∗
, Thomas Wick2 and Yan Jin1
1 State Key Laboratory of Petroleum Resources and Prospecting, China University of Petroleum, Beijing, China
2 Institut für Angewandte Mathematik, Leibniz Universität Hannover, Welfengarten 1, 30167 Hannover, Germany
Abstract. The goal of this work is to incorporate a directional strain energy splitting method into a monolithic
phase-field model for mixed-mode fracture propagation. Failure of rocks is always accompanied by the prop-
agation of mixed-mode fractures. Some recent studies have incorporated mixed-mode fracture propagation
criterion to phase-field models. But these new models have the shortcomings of being not physically sound
or not numerically stable. And they are all solved with staggered scheme. In this work, an existing energy
splitting method for masonry-like material is modified and incorporated into the phase-field model to account
for mixed-mode fractures. A fully-monolithic scheme is used to solve the model. A primal-dual active set
method, which is a semi-smooth Newton approach is utilized to treat fracture irreversibility. Mixed-mode
fracture tests are carried out, and the results of the new model are compared to those of existing models.
1 Introduction
In classic fracture mechanics, three fundamental modes of fractures are defined. They are termed as
tensile (mode-I), in-plane shear (mode-II) and anti-plane shear (mode-III) [2]. But the fracture we deal
with in reality is mostly mixed-mode. For example, mode-I-II fracture propagation is very common in
rocks under compression. A modified G-criterion for mode-I-II fracture propagation is proposed in [10].
In recent studies, this criterion is incorporated into the phase-field fracture model by splitting the crack
driving energy into mode-I and mode-II parts. More specifically, the crack driving stress σ + is splitted
into σI+ and σII
+
. In [13], the energy splitting method proposed in [9] is modified so that the first and
second terms of σ + correspond to σI+ and σII +
respectively. But this method overestimates the maximum
force response when mode-I loading is dominant. More physically-sound splitting methods are proposed
in [12], [11] and [3], where tension-compression asymmetry normal to the fracture surface is fulfilled
explicitly and the fracture surface is considered as frictionless. But in our experiments, these models are
numerically unstable. In [4], a splitting method for masonry-like material is derived based on structured
deformation theory under principal strain coordinates. This method has a similar formulation to [12]
and [11], but is more stable. However, it only accounts for mode-I fracture. Inspired by [3], the splitting
method proposed in [4] is modified in this work to account for mode-I-II fractures.
2 Governing equations for a mixed-mode phase-field fracture model
In [13], the modified G-criterion for mixed-mode fracture propagation is incorporated into the thermody-
namically consistent phase-field model proposed in [9]. The formulation reads: Find displacements u ∈ V
and a phase-field variable ϕ ∈ Win such that at each incremental step:

(1 − κ) ϕ2 + κ σI+ (u) + σII +
(u) , e (w) + σ − (u) , e (w) =0 for all admissible w, (1)
+ +
σI (u) σII (u) 1
(1 − κ) ϕ + : e (u) , ψ + (ϕ − 1, ψ) + (∇ϕ, ∇ψ) ≥0 for all admissible ψ. (2)
GIc GIIc
Here, σ and e denote stress and strain respectively. is the characteristic length scale of smeared fracture.
κ is a small numerical parameter to prevent singularity in Jacobian matrix. w and ψ are test functions
for u and ϕ. GIc and GIIc are mode-I and R mode-II critical energy release rates. Finally, we denote L
2
scalar product with (A, B) := (A, B)B := B A : Bdx

∗ Corresponding author: 2016312016@student.cup.edu.cn
51
3 Stress splitting for mixed-mode phase-field fracture model
In [13], σI+ and σII

+
are determined as the first and second terms of the σ + in [9]. This splitting method can
capture the wing crack geometry in uniaxial compression test by suppressing the early-stage propagation
of shear cracks, but it overestimates the maximum force response under tension-dominant loading.
In [4], a splitting method is derived based on structural strain theory under principal strain coordinate.
In 2D case, Let p1 denotes the direction of eigenvector corresponding to the most-positive eigenvalue.
And p2 denotes the direction of the other eigenvector. The method can be described in the strain form:

 ep1 0


 if ep2 > 0,

 0 ep2



+
e = ep1 + νep2 0
if ep2 ≤ 0 and ep1 + λ+2µλ
ep2 > 0, (3)

 0 0






 0 if e + λ e ≤ 0.
p1 λ+2µ p2
This splitting method can capture the tensile fracture geometry under compression loading [5]. But it
does not account for mixed-mode fractures.
Therefore, we do a heuristic modification to Freddi’s splitting method. More specifically, the reference
frame is changed from principal strain coordinate to local crack surface coordinate. In 2D case, let n and
s denote the directions normal and tangential to the crack surface. The method can be described as:

 e nn 0 λ


 if enn + λ+2µ ess > 0 and ess > 0,

 0 e ss



+
eI = enn + νess 0 λ
if enn + λ+2µ ess > 0 and ess ≤ 0, (4)

 0 0






 0 if e + λ e ≤ 0,
nn λ+2µ ss

1 0 ens
e+
II = . (5)
2 esn 0
As for the crack surface direction, we refer to [3] to solve an additional local dissipation maximization
problem. Let θ denote the angle between n and p1, we define the normalized crack driving energy as:
!
σI+ (u, θ) +
σII (u, θ)
F (u, θ) = + : e (u) (6)
GIc GIIc
then the normal direction of crack surface can be determined by:
θn = Arg{ maxπ F (u, θ)}. (7)
θ∈[0, 2 )
After simple derivation, we can get:



 2 1 2 λ

 µ (ep1 − ep2 ) − sin 4θ if enn + e
λ+2µ ss
> 0 and ess > 0,

 GIIc GIc





 2 1 1

 µ (ep1 − ep2 ) − sin 4θ

 GIIc GIc
∂F (u, θ)
= 2 2 (8)
∂θ 
 (λ + 2µ) e p1 − ep2
−
 sin 2θ if enn + λ
e > 0 and ess ≤ 0,

 λ+2µ ss

 G Ic



 2

 µ (ep1 − ep2 ) sin 4θ

 if enn + λ
e ≤ 0.
λ+2µ ss
GIIc
In all three situations, the local maximum points of F (u, θ) are located at θ = 0 or θ = π4 or both.
Therefore, instead of implementing complicated search algorithm, we can determine the value of θn by
comparing F (u, 0) and F (u, π4 ) and choose whichever maximize F (u, θ).
52
4 Numerical schemes
The nonlinear quasi-monolithic displacement/phase-field system presented in the Sections 2 and 3 is solved
with Newton’s method and line search algorithms. The constraint minimization problem is treated with
a semi-smooth Newton method (i.e., a primal-dual active set method). Both methods are combined in
one single loop leading to a robust and efficient iteration scheme that is outlined in [7]. Within Newton’s
loop we solve the linear equation systems with GMRES solvers with a diagonal block-preconditioning
AMG. The overall implementation is based on deal.II [1]; for details we refer to [6].
5 Numerical simulations
To demonstrate the capabilities of different energy splitting methods in capturing mixed-mode fracture
geometry, we use the experiment in [8] as benchmark and compare the results of our method with those
of Miehe’s [9], Zhang’s [13] and Freddi’s [4] methods.
The sample is a 104 mm × 80 mm rectangle marble plate with a notch at the center. The notch is
20 mm in length and 1 mm in width and is inclined at 45◦ to the horizontal plane. Young’s modulus
E = 63.5 GPa [8], Poisson’s ratio ν = 0.21, and plane stress setting is used. Mode-I critical energy
2 2
release rate GIc = 11 J/m and mode-II critical energy release rate GIIc = 110 J/m [13]. Two
uniform-refined meshes, with 7952 and 30272 cells respectively, are used in the simulation. The minimum
mesh sizes h are 1.17 mm and 0.63 mm in two cases. Characteristic length = 2h. Numerical parameter
κ = 10-10 . Newton solver tolerance is set as 10-6 . The sample is subjected to displacement-controlled
uniaxial compression loading at the top boundary with the time step size ∆t = 10-3 s. The crack
geometries using the fine mesh are shown in Figure 1. According to [8]. The crack geometry should
(a) Miehe’s splitting method (b) Zhang’s splitting method (c) Freddi’s splitting method (d) Our new splitting method
Figure 1. Simulated crack geometries using four splitting methods.
have two features: (a) tensile fractures sub-parallel to the loading direction, which are typical mode-I
fractures; (b) "X-shape" shear bands, which are typical mode-II fractures.
It is obvious that only our new splitting method can capture both features. The load-displacement curves
on the upper boundary is shown in Figure 2. All four splitting methods are numerically stable in both
meshes. The simulation continues until the load starts to decrease, when cracks reach the boundary.
And the load-displacement curves in two meshes agree well with each other in the early stage. However,
the maximum load response in the finer mesh is 25% - 30% higher than that in the coarser mesh. This
phenomenon has also been observed in [14].
(a) Miehe’s splitting method (b) Zhang’s splitting method (c) Freddi’s splitting method (d) Our new splitting method
Figure 2. Load-displacement curves using four splitting methods.
53
Acknowledgments
This work has been supported by the German Research Foundation, Priority Program 1748 (DFG SPP
1748) in the subproject Structure Preserving Adaptive Enriched Galerkin Methods for Pressure-Driven
3D Fracture Phase-Field Models with the project No. 392587580. The authors are also grateful for the
financial support from the program of China Scholarships Council (No. 201806440069) and National
Natural Science Foundation of China (No. 51490651).
References
[1] Daniel Arndt, Wolfgang Bangerth, Denis Davydov, Timo Heister, Luca Heltai, Martin Kronbichler,
Matthias Maier, Jean-Paul Pelteret, Bruno Turcksin, and David Wells. The deal. II library, version
8.5. Journal of Numerical Mathematics 25.3 (2017), pp. 137–145.
[2] Barry Kean Atkinson. Fracture mechanics of rock. Elsevier, 2015. isbn: 1483292746.
[3] Eric C. Bryant and WaiChing Sun. A mixed-mode phase field fracture model in anisotropic rocks
with consistent kinematics. Computer Methods in Applied Mechanics and Engineering 342 (2018),
pp. 561–584. issn: 00457825.
[4] Francesco Freddi and Gianni Royer-Carfagni. Regularized variational theories of fracture: A uni-
fied approach. Journal of the Mechanics and Physics of Solids 58.8 (2010), pp. 1154–1174. issn:
00225096.
[5] Francesco Freddi and Gianni Royer-Carfagni. Variational fracture mechanics to model compressive
splitting of masonry-like materials. Annals of Solid and Structural Mechanics 2.2-4 (2011), pp. 57–
67. issn: 1867-6936.
[6] Timo Heister and Thomas Wick. Parallel solution, adaptivity, computational convergence, and open-
source code of 2d and 3d pressurized phase-field fracture problems.
[7] Timo Heister, Mary F. Wheeler, and Thomas Wick. A primal-dual active set method and predictor-
corrector mesh adaptivity for computing fracture propagation using a phase-field approach. Com-
puter Methods in Applied Mechanics and Engineering 290 (2015), pp. 466–495. issn: 00457825.
[8] Huang Jiefan, Chen Ganglin, Zhao Yonghong, and Wang Ren. An experimental study of the strain
field development prior to failure of a marble plate under compression. Tectonophysics 175.1-3
(1990), pp. 269–284. issn: 00401951.
[9] Christian Miehe, Fabian Welschinger, and Martina Hofacker. Thermodynamically consistent phase-
field models of fracture: Variational principles and multi-field FE implementations. International
Journal for Numerical Methods in Engineering 83.10 (2010), pp. 1273–1311.
[10] B Shen and O Stephansson. Modification of the G-criterion for crack propagation subjected to
compression. Engineering Fracture Mechanics 47.2 (1994), pp. 177–189.
[11] Christian Steinke and Michael Kaliske. A phase-field crack model based on directional stress de-
composition. Computational Mechanics 55 (2018), p. 1017. issn: 0178-7675.
[12] Michael Strobl and Thomas Seelig. A novel treatment of crack boundary conditions in phase field
models of fracture. PAMM 15.1 (2015), pp. 155–156. issn: 16177061.
[13] Xue Zhang, Scott W. Sloan, Chet Vignes, and Daichao Sheng. A modification of the phase-field
model for mixed mode crack propagation in rock-like materials. Computer Methods in Applied
Mechanics and Engineering 322 (2017), pp. 123–136. issn: 00457825.
[14] Xue Zhang, Chet Vignes, Scott W. Sloan, and Daichao Sheng. Numerical evaluation of the phase-
field model for brittle fracture with emphasis on the length scale. Computational Mechanics 59.5
(2017), pp. 737–752. issn: 0178-7675.
54

Load Optimisation for Damage Reduction and Fracture

Prevention
1
Fabian Guhr ∗
and Franz-Joseph Barthold1
1 Structural Mechanics, TU Dortmund University, Dortmund, Germany
Abstract. In this paper, load optimisation is applied to air bending with the aim of reducing the damage
accumulation within the critical domain of the forming process, and as such preventing fracture. A framework
for load optimisation with the commercial software Abaqus is provided. Modelling the forces with polynomials
and the load increments with B-splines, respectively, allows an easy implementation of load optimisation.
Introduction
The cause for macroscopic fracture is the accumulation of damage within the material. Damage is
typically associated to voids and micro cracks at grain level. To prevent fracture at the macroscopic
scale, damage accumulation has to be controlled and reduced.
In [6], damage was investigated for different forming processes. For example in air bending, damage can
be reduced by superimposing compressive stresses in the critical area, i.e the area opposite where the tool
is in contact with the sheet. Due to the high triaxiality values in the outer boundary of the sheet, i.e.
a tensile dominated stress state. Damage accumulation in this region leads to the most critical point of
fracture during the forming process. To circumvent failure and reduce damage accumulation, the authors
of [6] developed the so-called process of elastomer bending. There, an elastomer cushion is placed into
the die, such that the resulting compressive strains lead to a reduced damage growth by reducing the
triaxiality, and thereby the stress state in the critical area.
To optimise the process numerically, the elastomer cushion, which would generate pressure forces as
reaction forces, is replaced with prescribed nodal forces at the corresponding boundary. By using mathe-
matical optimisation with the resulting external loads as the design, a new optimal load can be generated.
Furthermore, by optimising the load amplitudes for the nodal loads, the load application can further be
optimised to control the triaxiality during the forming process.
1 Theoretical framework
For the implementation of load optimisation, the external forces in the equilibrium condition within the
finite element (FE) formulation, i.e.
R = Fint − Fext = K · u − λf = 0, (1)
are controlled. This leads to the following two possible adaptations:

a) load optimisation of external forces f
b) load-path optimisation of the scaling factor for the numerical increments λ.
∗ Corresponding author: fabian.guhr@tu-dortmund.de ORCID iD: 0000-0002-3341-0503
55
The idea for optimisation of external loads (a) has been presented in [2] and [3]. There, the external
loads were prescribed by a polynomial function
f = an xn + an−1 xn−1 + · · · + a1 x + a0 , (2)
with the coefficients ai of the polynomial as the design variables within the optimisation.
To further optimise the process beyond the concept of elastomer bending, the load application over time
can also be optimised via the load scaling parameter λ (b). To model a more applicable and arbitrary
load-path, B-splines are chosen to define the scaling parameter over the course of the forming process.
B-splines are curves generated from multiple polynomials and defined by control points C i . As such, a
point b(t) = λ(t) on the B-spline at t ∈ [0, 1] is defined as
n
X
b(t) = C i Nid (t), (3)
i=0
with the basic functions Ni , the number of control points n and the degree of the curve d. Additionally,
a knot vector T has to be defined
T = {t0 , t1 , . . . , tm } , (4)
with m = d + n + 1. In addition, the first d + 1 knots need to be equal to zero, as well as the last p + 1
to be equal to one. The remaining ti have to fulfil the condition ti ≤ ti+1 . The basis functions Ni are
(
j 1, for t ∈ [ti , ti+1 ]
Ni (t) = (5)
0, else
when j = 0. Otherwise, the basis functions result in
t − ti ti+j+1 − t
Nij (t) = N j−1 (t) + N j−1 (t), (6)
ti+j − ti i ti+j+1 − ti+1 i+1
where j = 1, 2, . . . , d. Generally, the de-Boor algorithm is used for efficient calculation of the above basis
functions and therefore the B-spline. For more information regarding B-splines, c.f. [1].
2 Numerical implementation
The main aspect of this work is the implementation of a optimisation

framework for the above case with the commercial software Abaqus
.inp le
in mind. The two main advantages are the direct comparibility to
the results obtained in [6], as well as the contact algorithms within
Abaqus, which are a necessity for the forming process at hand. For
the simulation, an elasto-plastic model without damage mechanics is
Str used. Instead, the triaxiality is used to constrain the damage growth,
uct
ura
lr
c.f. [4, 5].
esp
ons
e For the framework, the python interface, which is provided by Abaqus,
is used to extract the data from the output data base (odb) as
well as transferring that data to Matlab for further processing. The
Gra mathematical optimisation, i.e. Sequentially Quadratic Programming
die
nts (SQP), is done within Matlab. This requires the additional informa-
tion of gradients for the objective function and constraints. These
Optimisation are generated by perturbing the design variables and calculating the
numerical gradients with additional FEM simulations.
To transfer data back to the Abaqus simulations, the input file is edited
Figure 1. The overall framework for
with the newly generated loads or amplitude information. For the ex-
the optimisation. ternal nodal loads, this requires to define so-called sets within Abaqus
to define the areas of interest, i.e. objective function, constraints and
the area where nodal forces are applied. With this defined, the nec-
essary information can be extracted and the exact nodal forces mapped to their respective nodes. For
the load-path optimisation, tabular amplitudes are chosen. By exchanging the predefined data with the
optimised B-spline data, new load-paths can be defined for the Abaqus simulation.
56
2.1 Optimisation problem
The main idea behind the optimisation definition is to generate a structure that is as close to the initial
shape as possible. This shape is captured with the initial deformation field Upre and compared to the
current shape U(s) generated by the optimisation. To minimise the difference between those two results,
the objective function J is chosen to be of least square type
minimise J(s) = kU(s) − Upre k22
s
(7)
subject to η(s) ≤ η max ,
where s the vector of design variables, i.e. either the coefficients of the polynomial ai for the load
optimisation, or the vector of controls points C for the load-path optimisation. To reduce the damage,
the triaxiality η is chosen as the constraint. The triaxiality η is defined as
σh
η= , (8)
σvM
where σvM are the von Mises stresses and
I1 tr(σ)
σh = = (9)
3 3
as the hydrostatic stresses. To constrain the triaxiality, and to compare the results to the ones obtained
in [6], the critical triaxiality η max is set to η max = 0.48.
The constraint for the optimisation problem in Eq. (7) is applied differently for the two types of op-
timisations at hand. In the case of external load optimisation (a), the triaxiality is only controlled at
the end of the simulation and as such, the triaxiality can exceed the threshold during the simulation.
To furthermore control the triaxiality during the process, the load-path optimisation (b) controls the
triaxiality at each load-step during the optimisation.
3 Results
The process to be optimised is depicted in Fig. 2. The

red area is used to apply the necessary external forces in
x- and y-direction, such that a resulting force in any of
the two directions is possible. This area corresponds to
the area where the elastomer cushion in [6] generates the
necessary reaction forces. Node A, depicted by the small
green dot, is later on used for evaluating the optimal load-
path. Currently, both optimisations are done separately,
as well as the load-path optimisation only supporting one
node at a time.
For the external load optimisation, a small constraint is
added, wherein the loads at point A shall remain per-
pendicular to the surface, which results in the condition
f x (x = A) = 0 → ax0 = 0. (10)
The results for the external load optimisation are pre-

Figure 2. The simulation and optimisation setup.
sented in Fig. 3. The left side shows the initial, unaltered
The deformable sheet is depicted in blue
and the loads are applied at the red deformation and the corresponding triaxiality state. Ap-
boundary. plying the optimised loads, the triaxiality can be reduced
below the threshold of η max in a certain area around the
applied forces. Compared to the results obtained in [6], the triaxiality around the area of interest shows
the same characteristic triaxiality reduction.
With these optimised loads at hand, the load-path is then optimised for node A, see Fig. 2, since the
triaxiality exceeds the imposed threshold due to the linear load-path during the forming process, see the
blue curve in Fig. 4. Optimisation of the scaling parameter λ generates the non-linear load-path which
then allows constraining the triaxiality at each time step and subsequently the critical value is never
exceeded, see the red curve in Fig. 4.
57
a) b)
Figure 3. The triaxiality distribution in the deformed sheet. The left graph shows the initial shape with no
additional loads, the right the sheet with the optimised loads.
Figure 4. Results from the load-path optimisation. The triaxiality over the punch displacements on the left and
the corresponding optimised load-path. The values are for node A of Fig. 2.
Acknowledgments
Funding of projects C05 within the Collaborative Research Centre TRR 188 — Projektnummer 278868966
"Damage controlled forming processes" by the Deutsche Forschungsgemeinschaft (DFG, German Research
Foundation) is gratefully acknowledged.
References
[1] Gerald Farin. 8 - B-Spline Curves. In: Curves and Surfaces for CAGD (Fifth Edition). Ed. by Gerald
Farin. Fifth Edition. The Morgan Kaufmann Series in Computer Graphics. San Francisco: Morgan
Kaufmann, 2002, pp. 119 –146.
[2] Jan Liedmann and Franz-Joseph Barthold. Sensitivity Analysis of Nonlinear Structural Response
regarding Geometry and External Loads. PAMM 18.1 (2018), e201800135.
[3] Jan Liedmann and Franz-Joseph Barthold. Sensitvity Analysis of elastoplastic Structural Response
regarding Geometry and External Loads. Proc. ECCM6/ECFD7 (2018), pp. 1308 –1319.
[4] Rickmer Meya, Christian Löbbe, and A. Erman Tekkaya. Stress State Control by a Novel Bending
Process and its Effect on Damage Evolution. Joint MSEC-NAMRC-Manufacturing USA 2 (2018).
[5] Christian C. Roth and Dirk Mohr. Ductile fracture experiments with locally proportional loading
histories. International Journal of Plasticity 79 (2016), pp. 328 –354. issn: 0749-6419.
[6] A. Erman Tekkaya, Nooman Ben Khalifa, Oliver Hering, Rickmer Meya, Sebastian Myslicki, and
Frank Walther. Forming-induced damage and its effects on product properties. CIRP Annals 66.1
(2017), pp. 281 –284. issn: 0007-8506.
58

A phase-field fracture model based on Mindlin’s strain

gradient theory
Resam Makvandi1 and Daniel Juhre ∗ 1
1 Institute
of Mechanics, Faculty of Mechanical Engineering, Otto von Guericke University Magdeburg, Universitätsplatz 2, 39106
Magdeburg, Germany.
Abstract. The classical theory of continuum mechanics utilized in the conventional phase-field fracture models
suffers from singular stress fields at the crack front. Although the so-called degradation function regularizes
the problem under consideration, the singular behavior is still present in the mathematical model and therefore
exerts a notable effect on the final results. Strain gradient models do not show the mentioned singular behavior.
Therefore, in this contribution, the integration of these models into the phase-field modelling of fracture is pro-
posed. It has been shown that the utilization of these models helps in mitigating the problem of stress singulari-
ties.
Introduction
The classical continuum mechanics theories are originally supposed to determine deformations in ranges
from millimeter to meter, the so-called macroscopic scales. In fact, these models are approximations
of physical systems neglecting the underlying microstructure [3]. For instance, the appearance of local
singularities at the crack tips (or more broadly, in the presence of point and line loads) is one of the
known limitations of the classical continuum mechanics theory. Generalizing these models by introducing
additional kinematic terms to consider the underlying microstructure effects at macroscopic levels is one
way of overcoming the already mentioned problems. Mindlin’s theory of elasticity with microstructures
[5] is one of the well-known works in this area. This broad-spectrum framework simplifies to strain
gradient elasticity theories by introducing the relation between the micro-deformations and gradients of
the macro-deformations. In gradient elasticity, the strain energy function contains both the strain terms
and their gradients, which leads to stresses depending on higher-order derivatives of the displacement
field. One of the motivations of proposing the gradient elasticity theory for solid mechanics is to overcome
the problem of singular fields in the vicinity of crack fronts and dislocation cores [4].
1 Governing equations
In this contribution, the combination of Mindlin’s based strain gradient elasticity models and phase-field
fracture mechanics is proposed. In this section, we will first derive the governing equations of the gradient
elasticity and then will use these equations in the phase-field modelling of fracture.
1.1 Strain gradient elasticity
In the following, only infinitesimal deformations of the continuum are assumed. Considering macroscopic
and microscopic deformations, based on the theory of elasticity with microstructures [5], the strain energy
density can be defined: (the contractions between tensors are defined as in [2])
1 h4i 1 h4i 1 h6i h5i h5i h4i

W = ε ·· C·· ε + γ ·· B·· γ + κ ··· A··· κ + κ ··· D·· γ + ε ·· F ··· κ + ε ·· G·· γ, (1)
2 2 2
∗ Corresponding author: daniel.juhre@ovgu.de ORCID iD: 0000-0001-8997-3818
59
h6i h4i h4i h5i h5i h4i

with A, B , C , D, F , and G being the constitutive tensors where the number in brackets defines the
order of the tensor. In Eq. (1), ε and γ denote the linear strain tensor and the relative deformation,
respectively, and are defined as
1 1
ε := ∇sym u = (∇ ⊗ u + u ⊗ ∇) = (ui,j + uj,i ) = εij , γ = (∇u − ϕ) , (2)
2 2
where u is the displacement field and ϕ represents the microscopic deformations. Moreover, we define
the gradient of macroscopic deformation as
κ := ∇∇u = ui,jk = κijk . (3)
Assuming linear elasticity and centro-symmetric isotropic materials, Eq. (1) reduces to
2
W̃ = a1 (ε ··I) + a2 ε ··ε + a3 (κ ··I) · (κ ··I) + a4 I ··κ · κ ··I + a5 (I ··κ) · (I ··κ) + a6 κ ··· κ + a7 κ ··· κ, (4)
where I is the identity tensor and a1 · · · a7 are (independent) material parameters. Equation (4) can be
re-written in index notation as [6]
1
W̃ = λεii εjj + µεij εij + b1 κiik κkjj + b2 κijj κikk + b3 κiik κjjk + b4 κijk κijk + b5 κijk κkji , (5)
2
where λ and µ are the Lamé constants and b1 · · · b5 are gradient material parameters. We define the
stress σ (energy conjugate to ε) and double stress τ (energy conjugate to κ) as
∂ W̃
σ := , σij := λεpp δij + 2µεij (6)
∂ε
∂ W̃ 1
τ := , τijk = b1 (δjk κppi + 2 δij κkpp + δik κppj ) + b2 (δjk κipp + δik κjpp )
∂κ 2 (7)
+ 2 b3 δij κppk + 2 b4 κijk + b5 (κkji + κkij )
To find the weak form, we now calculate the first variation of the work of the internal forces:
Z Z
int
δ W̃ = σ ·· δε dV + τ ··· δκ dV, (8)
Ω Ω
with δε := ∇sym (δu), and δκ := ∇∇(δu). Applying the divergence theorem one time on the first term
and two times on the second term of the the right-hand side gives us the necessary boundary terms,
Z Z Z
σ ·· δε dV = − (∇ · σ) · δu dV + (σ · n) · δu dA, (9)
Ω Ω Γ1
Z Z Z h i
T
τ ··· δκ dV = ∇2 τ · δu dV + − (∇ · τ ) · n + ∇s · n τ ·· [n ⊗ n] − ∇s · [τ · n] · δu dA
Ω
ZΩ Γ1
Z X (10)
+ [τ ·· [n ⊗ n]] · δ∇n u dA + [τ ·· [m ⊗ n]] · δu dL.
Γ1 Ξ e
In Eqs. (9) and (10), Ω represents the whole volume, Γ1 the surface boundary and Ξ the union of the
surface boundary intersection segments (a network of edges e). Moreover, n denotes the outer normal
to the surface, m is the outward normal vector to Ξ and is perpendicular to n. Finally, considering the
work of the external forces and the principle of virtual work, the weak form becomes
δ W̃ int − δ W̃ ext = 0,
Z Z Z Z Z Z
(11)
σ ·· δε dV + τ ··· δκ dV = b · δu dV + t̄1 · δu dA + t̄2 · δ∇n u dA, + t̄3 · δu dL.
Ω Ω Ω Γ1 Γ1 Ξ
where b, t̄1 , t̄2 and t̄3 are the vector of body forces, surface tractions, surface double tractions and edge
tractions, respectively.
60
1.2 Strain gradient enhanced phase-field fracture model
We will consider the conventional fourth-order phase-field fracture model. The total energy of fracture
can be represented by
Z Z
(1 − s)2 1 1 2
W̄frac (ε, κ, s) = s2 W̃ + Gc + |∇s|2 + 3 ∆s dV − b · u dV, (12)
Ω 4 2 4 Ω
where s is the phase-field parameter (s = 1 represents the intact material), Gc is the critical strain
energy release rate, represents the length-scale parameter for the crack, and g(s) = s2 is the so-
called degradation function. In Eq. (12), the strain energy density of the gradient material W̃ (see
Eqs. (4) and (5)) has replaced the classical linear elasticity strain energy density Welas = 21 λ(tr ε)2 +µε ·· ε.
Taking the first variations of Eq. (12) with respect to the displacement field and the phase-field parameter
gives us the weak form of the problem at hand. Introducing the effective stresses σeff = s2 σ and double
stresses τeff = s2 τ , and replacing them with the stress terms in Eq. (11) gives us the weak form of the
mechanical part. For the phase-field counterpart, the evolution of s is governed by
∂s δ W̄frac
= −M , (13)
∂t δs
where M ≥ 0 is the mobility constant. Equation (13) leads to the weak form for the phase-field
Z Z
1 ∂s 3 Gc (s − 1)
δs + Gc ∇s δ∇s + ∆s δ∆s + 2sW̃ + δs dV = qn δs dA, (14)
Ω M ∂t 2 2 Γ2
where qn is the normal flux.
2 Numerical simulations
The model derived above was implemented into the Finite Element Analysis Program (FEAP) [7] and
its isogeometric analysis add-on (igaFEAP). The domain with initial and boundary conditions are shown
in Fig. 1. We assume the plane strain state for all the simulations. The bottom edge of two-dimensional
body is fixed in the y-direction, while movements of the right edge in the x-direction are restricted. A line
displacement of magnitude 0.1 mm in the y-direction is applied to the top edge of the body in a quasi-static
manner. The initial crack length is 0.25 mm and is located in the middle of the plane, starting from the
left edge (see Fig. 1). The material properties used are λ = 50727.5 MPa, µ = 46825.4 MPa, Gc = 2.0 mm N
,
and M = 2000.0 MPa·s . For the gradient material properties, we use b1 = 0, b2 = 2 λl , b3 = 0, b4 =
1 1 2
2 µl , and b5 = 2 µl , where l represents the gradient length-scale parameter. This choice of gradient
1 2 1 2
material parameters is identical to using the GRADELA model developed by Aifantis and co-workers [1]
where the higher-order strains are defined as the gradients of the linear strain tensor. It should be noted
that setting l = 0 recovers the classical linear elasticity theory. Furthermore, we use the isogeometric
analysis and quadratic NURBS models with full integration as our model needs at least C 1 -continuous
elements. A mesh grid of 408 × 409 control points is used for the simulations shown in Fig. 2.
The results are shown in Fig. 2. The crack is expected to propagate along the path A-A0 . The results
are plotted when the maximum strain energy density in front of the crack tip reaches x = 0.65mm. Three
different values of l are considered. Behind the crack tip, the material is broken and consequently, the
existence of relatively high energy values (roughly 2.0 MPa) for the classical case (i.e. l = 0.0) does not
have a physical meaning. The gradient theory results (l > 0) does not show this behavior and the strain
energy density is completely zero behind the crack tip. Moreover, there is a clear difference between the
classical and the gradient theories in the results depicted for the path A-A0 . In the visualization of the
phase-field parameter, a wider transition zone between the cracked and the intact material is observed
for the classical theory.
Acknowledgments
The present work was financially supported by the Research Training Group 1554/2: Micro-Macro-
Interactions in Structured Media and Particle Systems of the DFG (German Research Foundation). This
support is gratefully acknowledged.
61
0.65 mm
initial crack
a) b)
Figure 1. a the initial crack position (defined by the phase-field parameter, i.e. non-geometric crack), dimensions
and boundary conditions of the problem, and b the phase-field state when the maximum strain energy
density in front of the crack tip reaches x = 0.65mm.
Figure 2. Results for the fourth-order formulation for different values of l and the 408 × 409 mesh . Distributions
of the phase-field (left), and the strain energy density (right) along three different cross-sections in the
plane.
References
[1] S. B. Altan and E. C. Aifantis. On the structure of the mode III crack-tip in gradient elasticity. Scr.
Metall Mater. 26.2 (1992), pp. 319–324. issn: 0956716X. doi: 10.1016/0956-716X(92)90194-J.
[2] A. Bertram. Compendium on Gradient Materials. 2017. doi: 10.13140/RG.2.2.36769.51045.
[3] F. dell’Isola, P. Seppecher, and A. Della Corte. The postulations á la D’Alembert and á la Cauchy for
higher gradient continuum theories are equivalent: a review of existing results. Proceedings. Math-
ematical, physical, and engineering sciences 471.2183 (2015), p. 20150415. issn: 1364-5021. doi:
10.1098/rspa.2015.0415.
[4] M. Lazar, G. A. Maugin, and E. C. Aifantis. Dislocations in second strain gradient elasticity. In-
ternational Journal of Solids and Structures 43.6 (2006), pp. 1787 –1817. issn: 0020-7683. doi:
10.1016/j.ijsolstr.2005.07.005.
[5] R. D. Mindlin. Micro-structure in linear elasticity. Arch. Ration. Mech. Anal. 16 (1964) (1964),
pp. 51–78.
[6] S.-A. Papanicolopulos and A. Zervos. Numerical solution of crack problems in gradient elasticity.
Proc. Inst. Civ. Eng.: Eng. Comput. Mech. 163.2 (2010), pp. 73–82. doi: 10.1680/eacm.2010.163.
2.73.
[7] R. L. Taylor. FEAP - Finite Element Analysis Program. 2014. url: http : / / projects . ce .
berkeley.edu/feap/.
62

Modelling of interfacial crack propagation in strongly

heterogeneous materials by using phase field method
1
Thanh-Tung Nguyen ∗
, Julien Yvonnet2 , Danièle Waldmann1 and Qi-Chang He2
1 University
of Luxembourg, Laboratory of Solid Structures, 6 Rue Richard Coudenhove-Kalergi, L-1359, Luxembourg
2 Université
Paris-Est, Laboratoire Modélisation et Simulation Multi Échelle MSME, UMR 8208 CNRS, 5 bd Descartes, F-77454
Marne-la-Vallée, France.
Abstract.
Phase field model has been proved to be a useful tool to study the fracture behaviors in heterogeneous materials.
This method is able to model complex, multiple crack fronts, and branching in both 2D/3D without ad-hoc
numerical treatments. In this study, a new interfacial cracking model in the phase field framework is proposed.
The effects of both stiff and soft interphases on the fracture response of composite materials are considered.
A dimensional-reduced model based on a rigorous asymptotic analysis is adapted to derive the null thickness
imperfect interface models from an original configuration containing thin interphase. The idea of mixing the
bulk and interfacial energy within the phase field framework is then used to describe the material degradation
both on the interface and in bulk. Moreover, in order to ensure the physical crack propagation patterns, a
unilateral contact condition is also proposed for the case of spring imperfect interface. The complex cracking
phenomena on interfaces such as initiation, delamination, coalescence, deflection, as well as the competition
between the interface and bulk cracking are successfully predicted by the present method. Concerning the
numerical aspect, the one-pass staggered algorithm is adapted, providing an extremely robust approach to
study interfacial cracking phenomena in a broad class of heterogeneous materials.
Introduction
Study of interface failure has been the topic of an intense research in the last decades. Cohesive element
[11] is a powerful method to model crack propagation along interface (situation where the crack path
is known). Other techniques like XFEM [4] have been also used to study interfacial cracking in bi-
materials [9], where the displacement jump can be accurately reproduced through the enrichment of
the displacement field with discontinuous functions. The XFEM technique has been used along with
the cohesive model to simulate delamination in composite materials [10]. However, XFEM methods can
hardly handle triple junctions or complex interphases, such as those in polycristals. More recently, the
interfacial cracking has been studied by using the phase field model [5].
This work is dedicated to developing a phase field framework for dealing with damage of a larger class
of imperfect interfaces, i.e. exhibiting either displacement jump or traction jump across an interface. In
order to avoid meshing the interphase, we employ an imperfect interface model of zero thickness (with
appropriate jump conditions), which is derived from a dimensional-reduced model based on a rigorous
asymptotic analysis. The jump conditions depend on the ratio of the stiffness of the interphase to those of
the connected materials and lead to two well-known models in the literature: coherent imperfect interface
for very stiff interphase, and linear spring-layer imperfect interface for very soft interphase [12, 13]. The
framework proposed in [5] is extended by including these models within the phase field framework, so as
to allow interaction between the damage in the thin interphase and the bulk cracks in an efficient and
robust manner.
∗ Corresponding author: thanh-tung.nguyen@uni.lu
63
The idea of mixing the bulk and interfacial energy within the phase field framework is used to describe
the material degradation both on the interface and in bulk [5]. It implies the total energy in a standard
framework of sharp discontinuity description as follows
Z Z Z Z
E= ψ e dΩ + ψ d dΓ + ψ ci dΓ + ψ si dΓ, (1)
Ω/Γd ,ΓI Γd ΓI,α ΓI,β
where ψ e is the elastic energy density; ψ d is the surface (fracture) energy density; ψ ci and ψ si are the
densities of interfacial energy of stiff interface, and compliant interface, respectively.
We adapt here the regularized description to both crack and interface. Herein, the sharp crack is de-
scribed by the smeared crack density function γd (d, ∇d), and the sharp interfaces are replaced by the
smeared interfaces, i.e., γα (α, ∇α) for stiff interphase and γβ (β, ∇β) for compliant interphase, respec-
tively. Furthermore, we substitute the displacement jump [[u]] by a smeared displacement jump v(x) (see
e.g. [5, 7] for more details). Then, the infinitesimal strain tensor ε in this framework can be decomposed
into a part related to the bulk εe and a part induced by the smoothed jump at the interfaces ε̃, implying
ε = εe + ε̃. The aforementioned energy functional can be now rewritten as
Z Z Z Z
E= e e
ψ (ε , d) hα,β dΩ + gc γd (d, ∇d) dΩ + ci s
ψ (ε , d)γα (α, ∇α) dΩ + ψ si (v, d)γβ (β, ∇β) dΩ,
Ω Ω Ω Ω
(2)
in which, hα,β = (1 − α)(1 − β) is introduced to distinguish the elastic energy of bulk phases from the
one of the interfaces; gc here denotes the fracture resistance; the tangential strain εs is expressed by†
εs = Tεe T = Mt εe , with Mt = T⊗T, and T(x) = I − n(x) ⊗ n(x), (3)

where n(x) is the unit normal vector field on the interface ΓI , and I is the second order unit tensor.
The failure model for both bulk and interfacial crackings is constructed through the unilateral contact
condition to maintain the physical crack propagation.
Bulk cracking
The model proposed by Miehe et al [3] with the assumption that damage induced by traction only is here
used for the bulk cracking
λ 2 2
ψ e (εe , d) = g(d) ψ e+ (εe ) + ψ e− (εe ) , with ψ e± (εe ) = htr εe i± + µ tr εe± , (4)
2
where g(d) = (1−d)2 + (with 1) is the degradation function, see [1, 3]); εe+ and εe− are, respectively
the extensive and compressive modes of the elastic strain tensor.
Interfacial cracking: coherent imperfect interface model
The interfacial cracking of the very stiff interface is assumed to be created by the whole tangential strains,
implying

1 s ci
ci ci s ci ci
ψ = g(d)ψe = ε : Cs (d) : ε , and Cs (d) = g(d) λs T ⊗ T + 2µs T⊗T . ci
(5)
2
The Lamé’s constants characterizing the interface, λci

s , and µs are given following [12].
ci
Interfacial cracking: spring imperfect interface model

The interfacial cracking of the very soft interface is assumed to be created by the tangential displacement
jump vt and the positive part of normal displacement jump vn+ . This leads to a decomposition expressed
by v = vn+ + vn− + vt , and ψ si is postulated as

1 1
ψ = g(d) ψn+ + ψt + ψn− = g(d) vn · Cs · vn + vt · Cs · vt + vn− · Csi
si si si si + si + si −
s · vn , (6)
2 2
† The

tensor product operation A⊗B is defined by (A⊗B)ijkl = Aik Bjl + Ail Bjk /2 for any two second-order tensors
A and B.
64
with the tangential interface stiffness Csi

s being defined following [13] as Cs = λs n ⊗ n + µs (I − n ⊗ n).
si si si
The Lamé’s constants characterizing the interface, λs , and µs , can be determined following [13].
si si
The variational approach to fracture as proposed in Bourdin, Francfort and Marigo [2, 1] and developed
in a convenient algorithmic setting by Miehe [3] is adopted here. It implies two problems: (i) phase field
problem, corresponding to a minimization of the total energy with respect to the displacement field u
and (ii) mechanical problem, corresponding to a minimization of the energy with respect to the scalar
damage field d. The details of numerical implementation can be found in [6, 8].
2 Numerical examples: two dimensional simulation of a microstructure
A rectangular plate containing several inclusions is considered under tensile loading. The dimensions of
the considered structure are L × B = 1 × 1.6 mm2 , while the inclusion diameters are taken to range from
d = 0.06 mm to d = 0.3 mm. The detailed geometry and boundary conditions are provided in Fig. 1 (a).
y
1.0 mm
x
o 1.6 mm
(a) Geometry/boundary conditions (b) Soft interphase (c) Stiff interphase
Figure 1. 2D simulation of a microstructure containing multiple inclusions.
Both situations of very stiff interface (CI) and very soft interface (SI) are considered. The material prop-
erties are given in the Table. 1. The tangential Lamé’s constants for the interface can be determined fol-
lowing [12, 13] with the interphase thickness h = 1 µm, we obtain λssi = 105 GPa/mm, µssi = 30 GPa/mm
and λsci = 6.857 GPa.mm, µsci = 8 GPa.mm.
The displacements are prescribed along the y−direction for upper edge (y = B) while the displacements
along x are free. On the lower edge (y = 0), the displacements along y are fixed to zero, while the
displacements along x are free. The monotonic displacement increments of ∆uy = 2 × 10−5 mm have
been prescribed via 1000 time steps. Plane strain condition is assumed.
Table 1. Material properties of the spring/coherent interface model.

Parameter Matrix Inclusion Soft interphase Stiff interphase Unit
λ 18 60 4.5 × 10−2 6 × 103 GPa
µ 12 32 3 × 10−2 4 × 103 GPa
gc 5 × 10−4 3 × 10−3 gcsi,n = 3.75 × 10−5 , gcsi,t = 3.18 × 10−5 5 × 10−3 [kN/mm]
The obtained results of crack propagation are shown in Fig. 1(b)(c). A strong impact of interfacial
properties on the fracture behavior of the heterogeneous material is obtained. We capture that the soft
interface provides a major interfacial cracking mode, while the stiff interface induces a main bulk cracking
behavior.
The very complex behavior of the interfacial cracking is reproduced by the present model. The compe-
tition and interaction between bulk cracking and interfacial cracking are successively simulated. More
interestingly, the proposed model can predict the poorly post-cracking properties of strongly bonded
interface observed in the experiment. This demonstrated the performance of the present computational
framework. It then constitutes a promising tool to evaluate the mechanical performance of complex
composite materials.
65
3 Conclusion
In this work, we introduce a robust interfacial cracking model in the phase field framework based on the
mixing of bulk energy with the energy of null-thickness interface. The proposed model is able to consider
the complex interfaces with various properties, i.e, very stiff interphase and very compliant interphase.
We apply the present model to study the interfacial cracking in composite materials and to understand the
effects of thin interphase properties on the global behavior of advanced material and structures. The new
model predict very well the complex cracking phenomena on interfaces such as initiation, delamination,
coalescence, and deflection.
Acknowledgments
This research is supported by the project SeRaMCo within the programme INTERREG – NWE.
References
[1] B. Bourdin, G.A. Francfort, and J.J. Marigo. “Numerical experiments in revisited brittle fracture”.
Journal of the Mechanics and Physics of Solids 48.4 (2000), pp. 797–826.
[2] G.A. Francfort and J.J. Marigo. “Revisiting brittle fracture as an energy minimization problem”.
Journal of the Mechanics and Physics of Solids 46.8 (1998), pp. 1319–1342.
[3] C. Miehe, M. Hofacker, and F. Welschinger. “A phase field model for rate-independent crack propa-
gation: Robust algorithmic implementation based on operator splits”. Computer Methods in Applied
Mechanics and Engineering 199 (2010), pp. 2765–2778.
[4] N. Moës, J. Dolbow, and T. Belytschko. “A finite element method for crack growth without remesh-
ing.” International Journal for Numerical Methods in Engineering. 46(1) (1999), pp. 131–156.
[5] T.T. Nguyen et al. “A phase-field method for computational modeling of interfacial damage interact-
ing with crack propagation in realistic microstructures obtained by microtomography”. Computer
Methods in Applied Mechanics and Engineering 312 (2016), pp. 567 –595. issn: 0045-7825.
[6] T.T. Nguyen et al. “A phase field method to simulate crack nucleation and propagation in strongly
heterogeneous materials from direct imaging of their microstructure”. Engineering Fracture Me-
chanics 139 (2015), pp. 18–39.
[7] T.T. Nguyen et al. “Multi-phase-field modeling of anisotropic crack propagation for polycrystalline
materials”. Computational Mechanics 60 (2017), pp. 289–314.
[8] T.T. Nguyen et al. “Phase field simulation of interfacial cracking in heterogeneous materials”. Sub-
mitted (2018).
[9] N. Sukumar et al. “Partition of unity enrichment for bimaterial interface cracks”. International
Journal for Numerical Methods in Engineering 59.8 (2004), pp. 1075–1102.
[10] Y. Wang and H. Waisman. “Progressive delamination analysis of composite materials using XFEM
and a discrete damage zone model”. Computational Mechanics 55.1 (2015), pp. 1–26.
[11] X.P. Xu and A. Needleman. “Numerical simulations of fast crack growth in brittle solids”. Journal
of the Mechanics and Physics of Solids 42.9 (1994), pp. 1397–1434.
[12] J. Yvonnet, H. Le quang, and Q-C. He. “An XFEM/level set approach to modelling surface/interface
effects and to computing the size-dependent effective properties of nanocomposites”. Computational
Mechanics 42.1 (2008), pp. 119–131.
[13] Q-Z. Zhu et al. “Three-dimensional numerical modelling by XFEM of spring-layer imperfect curved
interfaces with applications to linearly elastic composite materials”. International Journal for Nu-
merical Methods in Engineering 88.4 (2011), pp. 307–328.
66

Improved Durability Prediction of Cord-rubber

Composites by the Material Force Method
∗
Carmen M. Popa, Clemens Gebhardt and Michael Kaliske
Institute for Structural Analysis, Technische Universität Dresden, 01062 Dresden, Germany
Abstract. The research at hand deals with durability analysis of cord-reinforced elastomers using the material
force concept. Different types of steel cord reinforcement are used, in order to determine the influence of the
cord construction on the fracture sensitivity of the composite product. The composite is analysed using the
Extended Finite Element Method to reduce the effort of the meshing process. Material forces are computed
at the crack tip area and a qualitative ranking of the different reinforcement types is obtained. The results
are compared to experimental investigations.
Introduction
Cord-reinforced elastomers are composite structures widely used in the production of industrial rubber
products, from conveyor and power transmission belts used in industrial plants and factories, to multilayer
high-pressure hoses and coated fabrics, as well as tyres, see [1]. Cord-reinforced elastomers are made of a
low-modulus rubber matrix − characterised by high deformation behaviour, twisted reinforcement − with
much higher strength and the adhesive film (dip), that creates a strong bond between the reinforcement
and the elastomer matrix, see [2]. The operational safety of industrial products is directly influenced by
the choice of reinforcement, specifically the material used, the size and configuration of each cord and
density of cords in the rubber product. Ideally, the bonding between the reinforcement and the elastomer
is perfect, as reinforcing cords are coated with various chemical compounds. Nevertheless, the cords are
cut in order to be mounted in the respective component, which leads to a crack-prone area appearing
at the end of each cord. During the lifetime of the cord-reinforced elastomer, this area has a higher
probability of developing micro-cracks that lead to mechanical failure of the product. The influence of
the reinforcement geometry on this behaviour is an important information for developing new types of
cords and ensuring increased durability of cord-reinforced elastomers.
In this contribution, the material force method is used to analyse the influence of four different types
of cord geometries on the fracture sensitivity of the composite product. For general applications of
configurational forces, the books by Gurtin [4] and Maugin [6] introduce comprehensive knowledge, while
general applications for a finite element implementation of material forces are presented by Müller et al.
[7], Näser et al. [8] and Steinmann et al. [10].
The reinforcement model is generated by a straightforward numerical algorithm that requires a small
number of input parameters to provide the geometry of the four twisted cords used in this work. In
addition, the complex meshing process to match the geometry of the twisted cords, used by the Finite
Element Method (FEM) is rendered unnecessary with the eXtended Finite Element Method (XFEM).
1.1 Constitutive material models
The elastomer matrix has a highly nonlinear stress-strain response to mechanical loading, displaying an
upturn of the stress-strain diagram at large strains. The elastomer matrix of the composite is described
67
by the well-known Yeoh hyperelastic constitutive model, a phenomenological formulation that appears
as a practical solution to the need of characterising the elastic properties of rubber.
The role of reinforcement in most industrial rubber products is to ensure a high tensile strength-to-weight
ratio, the required flexibility as well as low elongation and high mechanical impact resistance. In tyres,
the reinforcement defines significantly the shape, supports loads and provides the needed rigidity of the
structure for acceleration, braking or cornering. In this work, cord reinforcement is defined as long cords
of twisted filaments made of steel, that provide extra strength to the rubber composites where they are
embedded in. At large strain, cord materials (steel, nylon, polyester or fiberglass) exhibit nonlinear elastic
behaviour, which may be described adequately by a variant of the compressible Neo-Hooke approach,
according to Simo & Hughes, see [9]. The Neo-Hooke material model has the simplest formulation of a
hyperelastic model and provides good results in modelling the proposed cord reinforcement material.
1.2 Modelling complex geometries
According to Khoei, see [5], the FEM is not the best choice of analysis in case of structures with dis-
continuities, singularities or boundary layers. The disadvantage of the method comes from the need to
conform the discretisation to all the geometrical details of the modelled structure. The XFEM is a parti-
tion of unity method (PUM) that adds enrichment functions to the conventional finite element function
space, in order to capture discontinuous behaviour in an internal area of a body. What makes the dis-
tinction between the XFEM and the conventional FEM is XFEM’s capability of describing an internal
interface without the need of conforming elements at its position, which allows the method to spare the
time-consuming step of mesh generation in case of complex geometries. The following figure details how
enrichement functions determine the steel material of twisted cords embedded in an elastomer sample.
Figure 1. Enriched nodes in cord-rubber composite
The position of the cord-rubber interface is obtained through a reinforcement generating method. The
numerical approach is implemented in the XFEM framework; it defines a flexible number of 2D surfaces
that are extruded along a curvilinear direction which relates to the twist trajectory of the filaments. A
3D body is generated from a small number of user parameters that define the cord reinforcement. The
algorithm can apply different twist trajectories to each of the specified 2D surfaces, a case mostly used in
reality when modelling cord reinforcement. This proves the versatility of the numerical approach. Two
of the four cord reinforcement types analysed are shown in the following figure.
C D
Figure 2. Cords C and D − cross-section and 3D geometries
68
1.3 Fracture sensitivity analysis
In case of structures with existing flaws, voids, inclusions or other singularities, fracture mechanics is an
appropriate approach to determine the fracture sensitivity of the analysed structures.
In the material space, a solid is assumed to be inhomogeneous, anisotropic and possibly changing in time.
Following Eshelby’s thought experiment from 1951 [3], the computation of the material force on a defect
p in an elastic medium can be derived by using the energy-momentum tensor. The material force, also
called configurational force − the thermodynamic force that drives the change in position of the defect
− is defined arbitrarily as the negative gradient of the change in energy. Material forces are a vectorial
nodal output. At every nodal point of the finite elements chosen, a three-dimensional vector is plotted.
Its norm represents the value of the energy release rate and its vector predicts the inverse direction of
crack propagation. This is one of the main advantages the material force method has over the classical
J-integral method.
2 Application
The simulation setup is chosen to replicate the initial experimental setup, where sets of four cord rein-
forcement samples are embedded in a rubber matrix and undergo monotonic and cyclic tensile (pull) tests.
The details of the experimental setup are shown in Figure 3 and the resulting maximum displacement
and force are detailed in Table 1 below.
cord crit. force [N] max. disp [mm]

A 1073 145
B 1013 112
C 1128 126
D 1057 109
Figure 3. Experimental sample Table 1. Force-displacement values
The modelling strategy for such large-strain static tension tests is formulated by a submodelling technique.
In this work, the concept of submodelling is used to describe the crack area at the end of cord reinforcement
inside the model. Static simulations are performed on samples of rubber embedding the four different
types of steel cord reinforcement. The displacement values of the global sample are interpolated onto the
boundary of the local model and act as input parameters for the subsequent local simulation.
Figure 4. Material forces on crack tip area of composite Figure 5. Energy release rate at crack tip area − XFEM
element comparison between four cord types
Once the local analysis is performed, the material force post-processor is applied to the crack tip area.
The output of the post-processor are material force vectors, that indicate the crack propagation direction
69
as the opposite direction to the displayed vectors and the energy release rate. Figure 4 shows the material
force vectors at the crack area for cord B and Figure 5 details the energy release rate of all four composite
samples.
It can be seen that the material force vectors exhibit larger values on the left side of the crack tip area,
indicating a higher sensitivity of crack initiation along the direction of the reinforcement, at a slight angle.
The integration of the vectorial quantities over the crack area determines energy measures, which give
information on the influence of the reinforcement type on the overall fracture behaviour of the composite.
3 Conclusion
The present work proposes a methodology for the fracture sensitivity analysis of four different cord-
reinforced composites, using the XFEM combined with a cord generation technique and a post-processor
based on the material force concept. Static analyses are carried out with local models, in order to deter-
mine the impact the type of cord reinforcement has on the overall fracture behaviour of the composite.
The numerical results match the experiments, making the array of presented methods a good means of
analysing the proposed problem.
Acknowledgments
This work has been funded by the European Social Fund (ESF) and the Free State of Saxony, which is
gratefully acknowledged.
References
[1] G. Burrowes and B. Rodgers. Reinforcing Materials in Rubber Products. USA: Nokian Tyres, 2005.
[2] M. G. Pastore Carbone. “Investigating Mechanical Behavior of Cord-rubber Composites by Multi-
Scale Experimental and Theoretical Approach”. PhD thesis. Universita degli Studi di Napoli Fed-
erico II, 2011.
[3] J.D. Eshelby. The force on an elastic singularity. Philosophical Transactions of the Royal Society
of London 244 (1951), pp. 87–112.
[4] M. E. Gurtin. Configurational forces as basic concepts of continuum physics. Springer, New York,
USA, 2000.
[5] Amir R. Khoei. Extended Finite Element Method. Theory and Applications. Wiley, West Sussex,
UK, 2015.
[6] G. A. Maugin. Configurational forces: thermomechanics, physics, mathematics and numerics. Taylor
& Francis, West Sussex, UK, 2010.
[7] R. Mueller and G. A. Maugin. On material forces and finite element discretizations. Computational
Mechanics 29 (2002), pp. 52–60.
[8] B. Naeser, M. Kaliske, and R. Mueller. Material forces for inelastic models at large strains: appli-
cation to fracture mechanics. Computational Mechanics 40 (2007), pp. 1005–1013.
[9] J.C. Simo and T.J.R. Hughes. Plasticity, Viscoplasticity and Viscoelasticity - Computational As-
pects. Springer, Berlin, Germany, 1992.
[10] P. Steinmann. Application of material forces to hyperelastostatic fracture mechanics. Journal of
Mechanics and Physics of Solids 37 (2000), pp. 7371–7391.
70

Thermo-mechanical phase-field model for porous-ductile frac-

ture
1
J. Schulte * , M. Dittmann1 , F. Aldakheel2 and C. Hesch1
1 Chair of Computational Mechanics, University of Siegen,Paul-Bonatz-Strasse 9-11, 57076 Siegen, Germany
2 Institute of Continuum Mechanics, Leibnitz University Hannover, Appelstr. 11, 30167 Hannover, Germany
Abstract.
Phase-field methods to regularize sharp interfaces represent a well established technique nowadays. In fracture
mechanics, recent works have shown the capability of the method for brittle as well as ductile problems
formulated within the fully non-linear regime [4, 5].
In this contribution, we propose a novel framework to simulate porous-ductile fracture in isotropic thermo-
elasto-plastic solids undergoing large deformations [3]. Therefore, a modified Gurson-Tvergaard-Needelman
GTN-type plasticity model is combined with a phase-field fracture approach to account for a temperature-
dependent growth of voids on micro-scale followed by crack initiation and propagation on macro-scale. The
multi-physical formulation is completed by the incorporation of an energy transfer into thermal field such that
on the other hand the temperature distribution depends on the evolution of the plastic strain and the crack
phase-field.
Eventually, a number of numerical investigations show not only the possibilities of the approach for a multi-
physical analysis of complex material behavior, but also the accordance with experimental results in terms of
hardening, necking, crack initiation and propagation. Moreover, a further example based on the third Sandia
Fracture Challenge is applied to demonstrate the capability of the model for the prediction of three-dimensional
fracture pattern in complex geometries.
Introduction
The analysis of crack initiation and propagation in ductile materials plays an important role in predicting
failure mechanisms for various engineering applications. The phase-field approach to fracture has been
proven to be a very powerful technique to simulate crack phenomena in multi-physical environments [4,
5]. The goal of this work is to present a theoretical and a computationally efficient framework for ductile,
porous materials undergoing thermomechanical loading conditions in order to study the influence of the
growth of micro-voids, as well as the final rupture at the macro-scale [1, 2, 3].
A theory for coupled thermomechanical response at fracture undergoing large deformations is outlined in
this section. The underlying, three-dimensional (d = 3) system leads to a multi-field setting with seven
fields
{ϕ, s, θ, α, rp , F p , f } , (1)
which are the deformation map, the crack phase-field, the absolute temperature and the four plastic
fields, which are the equivalent plastic strain, its dual hardening force, the plastic deformation gradient
and the void volume fraction. Here, the void volume fraction is defined as
1 − f0
f =1− (2)
Jp
∗ Corresponding author: jonathan.schulte@uni-siegen.de
71
and is assumed as a micro-mechanically motivated damage variable, related to the plastic deformation
gradient through its determinant J p . f0 is the initial void volume fraction.
The deformation gradient F = F e F p and its determinant J = J e J p are decomposed multiplicatively
into elastic and plastic parts. Moreover, the fracture insensitive isochoric and volumetric parts are given
by (
Xh ig(s) (J e )g(s) if J e > 1
˜ e
F̄ = e −1/d e
(J ) λa na ⊗ N a and J˜ = e
, (3)
a
Je else
postulating that fracture requires a local state of tensile/shear deformation. λea are the elastic principal
stretches, g(s) is a polynomial degradation function and na and N a are the principal directions of the
left and right stretch tensors.
1.1 Energetic response function
b for the coupled problem takes the form

The stored energy function Ψ
b =Ψ
Ψ b e (F , F p , s, θ) + Ψ
b θ (θ) + Ψ
b p (α, ∇α, θ) + Ψ
b f (s, ∇s, ∆s), (4)
which reflects a coupling of gradient thermoplasticity with gradient damage mechanics. The isotropic
elastic contribution is decomposed
be = Ψ
Ψ b evol (J˜e (J, J p , s), θ) + Ψ ˜ e (F , F p , J, s)),
b edev (F̄ (5)
into volumetric and deviatoric parts. As a representative constitutive law used for the numerical examples
we introduce the thermomechanically extended Neo-Hookean material model
!
κ ( ˜e )2 − 1
J 3 1
b e ˜e
Ψvol (J , θ) = ˜e
− ln[J ] − βκ(θ − θ0 ) J − ˜e
(6)
2 2 2 J˜e
and
˜ e ) = µ (F̄
b edev (F̄
Ψ ˜ e − 3).
˜ e : F̄ (7)
2
Therein, µ > 0 and κ > 0 denote the shear modulus and the bulk modulus, respectively. θ0 is a reference
temperature and β is the linear thermal expansion coefficient.
The purely thermal contribution to the stored energy (4) is assumed to have the simple form

b θ θ
Ψ (θ) = c θ − θ0 − θ ln , (8)
θ0
where c ≥ 0 is a constant parameter representing the specific heat capacity. Moreover, the Piola-Kirchhoff
heat flux vector can be defined as
Q(F , s, θ, ∇θ) := −K(F , s, θ)∇θ (9)
to account for the heat transfer. This is known as Duhamel’s law of heat conduction, where K is the
material thermal conductivity tensor
K(F , s, θ) := [K0 (1 − wK (θ − θ0 ))(1 − s) + K conv s]C −1 . (10)
Note that in case of fracture, the conduction degenerates locally such that we achieve a pure convection
problem and the heat transfer depends on the crack opening width. Here, we formulate the conductivity
tensor K in terms of the phase-field parameter s. Moreover, wK is a thermal softening parameter, K0
is a conductivity parameter related to the reference temperature, K conv is a convection parameter and
C = F T F denotes the right Cauchy-Green tensor.
We focus on the equivalent plastic strain α and its gradient ∇α with the particular form for the energetic
contributions
Zα
lp2
b
Ψ (α, ∇α, θ) = yb(α̃, θ) dα̃ + y0 (θ) k∇αk2 .
p
(11)
2
0
72
Here, lp is a plastic length scale related to a strain-gradient hardening effect and accounts for size effects
to overcome the nonphysical mesh sensitivity of the localized plastic deformation in softening materials.
Moreover, yb(α, θ) is a temperature dependent isotropic local hardening function obtained form experi-
mental data.
b p with respect to α yields
The variational derivative of Ψ
b p = ∂α Ψ
rp := δα Ψ b p − Div[∂∇α Ψ
b p] (12)
reflecting the characteristics of the gradient-extended model under consideration.
Eventually, the phase-field fracture contribution has to be formulated. Therefore, the sharp-crack surface
topology Γ is replaced by a regularized functional
Z
1 2 lf l3
Γl (s) = b(s, ∇s, ∆s) dV with γ
γ b(s, ∇s, ∆s) = s + k∇sk2 + f (∆s)2 . (13)
B0 4lf 2 4
The functional is based on the crack surface density function γ b per unit volume of the solid and the
fracture length scale parameter lf that governs the regularization. Note that in the limit lf → 0, the
regularized crack surface functional Γl (s) converges to the sharp crack surface Γ. For ductile fracture, we
require additionally that lp ≥ lf . For the given fracture surface functional introduced in (13), we define
the critical energy required to create a diffusive fracture topology by
Z
W cr = gbc (α) γ b(s, ∇s, ∆s) dV, (14)
B0
in terms of the Griffith-type critical energy release rate gbc , which is decomposed additively into elastic
and plastic contributions gbc (α) = gc,p + gc,e exp[−ωf α], using the modeling parameters {gc,e , gc,p , ωf }.
Summarized, the phase-field fracture contribution is given in terms of the crack-density function as
b f (s, ∇s, ∆s) = gbc (α) γ
Ψ b(s, ∇s, ∆s)
gbc (α) 2 gbc (α) lf gbc (α) lf3 (15)
= s + ∇s · ∇s + (∆s)2
4lf 2 4
which defines the crack resistance force via the variational derivative with respect to s
b f = ∂s Ψ
rf := δs Ψ b f − Div[∂∇s Ψ
b f ] + ∆[∂∆s Ψ
b f ]. (16)
1.2 Dissipative response function
Regarding the plastic material behavior, we postulate a modified Gurson model for porous plasticity as

σ2
b p (τ , rp ) = eq + 2q1 f cosh 3 q2 p − 1 + (q1 f )2
Φ (17)
r p 2 2 r p
in terms of the Kirchhoff stress which

p is related to the Cauchy stress by σ = τ /J and the dissipative
resistance force rp . Here, σeq = 3/2 kτ dev /Jk represents the von Mises equivalent stress, whereas
p = 13 tr[τ /J] denotes the local pressure along with the growth-based void volume fraction f and fitting
parameters q1 ≈ 1.5 and q2 ≈ 1.0.
b f based on the constitutive representation
On the fracture part, we define a crack threshold function Φ
b f (H − rf ) = H − rf ,
Φ (18)
where the energetic driving force H is bounded by the crack resistance force rf dual to the fracture
phase-field s.
bp bp
The associated plastic evolution equations are defined as dp = λp ∂∂τ
Φ
and α̇ = −λp ∂∂rΦp , whereas the
bf
evolution equation for the crack phase-field reads ṡ = λf ∂(H−r
∂Φ
f).
Eventually, we define the internal dissipation density function Dint := νp τ : dp + νf Hṡ, where νp is a
constant dissipation factor typically chosen in the range of 85% to 95% in the context of thermoplasticity.
In addition, νf is introduced as fracture dissipation factor based on the discussions related to an energy
transfer into the thermal field in [6] and the references therein.
73
2 Numerical Example
To illustrate the performance and applicability of the thermo-porous ductile fracture model to complex
three-dimensional geometries, a numerical example is considered in this section. Therefore, we adopt
CAD data from the third Sandia Fracture Challenge and apply the steel like material setting given.
Within the quasi-static simulation setting, the lower end of the hollow body is fixed in space while the
upper end is moved upwards. No heat in- or outflow is allowed. Figure 1 shows the von Misses residual
stress distribution within the deformed configuration on fully ruptured state, where the fractured regions
with g < 0.15 are removed from the contour plot for visualization purposes.
Figure 1. Third SFC. Von Mises stress distribution at fully ruptured state.
Acknowledgments
Support for the present research was provided by the Deutsche Forschungsgemeinschaft (DFG) under
grant HE5943/8-1 and DI2306/1-1.
References
[1] F. Aldakheel. Micromorphic approach for gradient-extended thermo-elastic–plastic solids in the log-
arithmic strain space. Continuum Mechanics and Thermodynamics 29 (2017), 1207––1217.
[2] F. Aldakheel, P. Wriggers, and C. Miehe. A modified Gurson-type plasticity model at finite strains:
formulation, numerical analysis and phase-field coupling. Computational Mechanics 62 (2017), 815––
833.
[3] M. Dittmann, F. Aldakheel, J. Schulte, M. Krüger, P. Wriggers, and C. Hesch. Phase-field modeling
of porous-ductile fracture in finite thermo-elasto-plastic solids. submitted (2018).
[4] M. Dittmann, F. Aldakheel, J. Schulte, P. Wriggers, and C. Hesch. Variational Phase-Field Formula-
tion of Non-Linear Ductile Fracture. Computer Methods in Applied Mechanics and Engineering 342
(2018.), pp. 71–94.
[5] M. Dittmann, M. Krüger, F. Schmidt, S. Schuß, and C. Hesch. Variational modeling of thermome-
chanical fracture and anisotropic frictional mortar contact problems with adhesion. Computational
Mechanics 63 (2019), pp. 571–591.
[6] J. Schulte, M. Krüger, M. Dittmann, and C. Hesch. Multi-field modeling of thermomechanical coupled
fracture problems. Proceedings in Applied Mathematics and Mechanics (2018). doi: 10.1002/pamm.
201800348.
74

Thermo-Mechanical Coupling for Internal Thickness Ex-

trapolation Elements
1
Felix Töller ∗
, Stefan Löhnert2 and Peter Wriggers1
1 Institute of Continuum Mechanics, Appelstraße 11, 30167 Hannover
2 Institute of Mechanics and Shell Structures, August-Bebel-Straße 30, 01219 Dresden
Abstract.
Internal Thickness Extrapolation (InTEx) is a new finite element technology that enables the combination of
initially flat elements with bulk material models [11]. Usually flat elements like Cohesive Zone Elements use
traction separation laws as bulk material models need a strain like deformation measure. Such a deformation
measure cannot be determined in all directions without further assumptions by an element with an initial
thickness of zero.
Tailored Forming is a manufacturing process for hybrid material components [2]. Internal Thickness Ex-
trapolation elements can be used to model the joining zone in simulations. The forming process is usually
performed at elevated temperature; often the different raw materials of the hybrid components are intended
to be heated differently to obtain similar yielding properties. Hence, taking heat transfer through Internal
Thickness Extrapolation elements into account is essential for the simulation of joining zones during Tailored
Forming.
Introduction
Tailored Forming is a manufacturing process for hybrid material components [2], where firmly bonded
hybrid work pieces are formed. The joining zone is defined as the bonded region (via intermetallic com-
pounds and metallic bonds) and its surrounding that is directly influencing the bonding characteristics.
The thickness of the joining zone is in the order of a few micrometers [11]. Modelling of the joining zone
for macroscopic simulations is preferably done with initially flat elements due to its small thickness [4,
12]. Cohesive Zone Elements are initially flat and therefore often used for simulations of joining zones.
Though Cohesive Zone Elements necessitate special material models, called traction separation laws.
These cannot capture all relevant deformation and damaging modes during Tailored Forming. Therefore
Internal Thickness Extrapolation (InTEx) elements are applied [11]. This new finite element technology
allows for the combination of initially flat elements with bulk material models and thus circumvents the
disadvantages of traction separation laws.
1 Internal Thickness Extrapolation
A flat element geometry is a simplification of the thin physical geometry of the joining zone. The basic idea
of Internal Thickness Extrapolation is to revert this simplification to reconstruct the physical geometry
of the joining zone. The thickened geometry can incorporate bulk material models. Figure 1 shows
the discretised geometry and the extrapolated (InTEx) geometry; for a better visibility the discretised
geometry here is not flat.
∗ Corresponding author: toeller@ikm.uni-hannover.de ORCID iD: 0000-0002-7457-2648
75
u
xI,InTEx
κI
u
xI,discr
l
xI,discr
-κI
l
xI,InTEx
Figure 1. Discretised and extrapolated geometry in the current configuration [10]
2 Heat Transfer
The heat balance equation is [8, 13]

∂Θ
div(q) + ρ cp =f . (1)
∂t
Here no internal heat source (f = 0) will be considered and the thermal state is locally assumed to be
stationary ( ∂Θ
∂t = 0). So the weak form of heat conduction is
Z Z
T
∂Θ q · n dA + k gradT (Θ) · grad(∂Θ) dV = 0 . (2)
∂Ω Ω | {z }
=gradT (∂Θ)·grad(Θ)
P
The discretised form can be obtained using ∂Θ ≈ ∂Θh = NI ∂ΘI and Θ ≈ Θh = NI ΘI where I is
omitted for simplicity
Z Z
∂ΘI NI q T · n dA + ∂ΘI k gradT (NI ) · grad (NJ ) dV ΘJ = 0 . (3)
∂Ω Ω
| {z }
sIJ
So the stiffness matrix is

 
Z Z NJ,x Z

sIJ = k gradT (NI ) · grad (NJ ) dV = k NI,x NI,y NI,z · NJ,y  dV = k BIT · BJ dV .
Ω Ω NJ,z Ω
(4)
As no internal heat source is considered, the heat flux is the residuum
q h = s · Θh . (5)
The thermal conductivity k to be used in the joining zone has to be determined. It is assumed to depend
on the damage in the joining zone. The thermal behaviour of a hybrid tube is investigated in [5] (Figure 2
left). It is intended to form a bearing bushing using the tube (Figure 2 right). For the forming process
the materials are heated differently to obtain similar yielding properties; i.e. the temperature of steel
(inside) has to be higher than the temperature of aluminium (outside) [3]. In the following the evolution
of the temperature distribution is simulated with a special focus on the heat transfer through the joining
zone. The initial temperatures are chosen according to the status “after heating” in [5] (initial in Figure
3). The transportation time (6 s from heating to the forming machine) is simulated and the temperatures
are compared to the results in [5] (InTEx and Reference in Figure 3). The work piece is only press fitted
and not firmly bonded which is here assumed to be equivalent to a fully damaged state (D = 1) of the
joining zone material.
For the simulations the following material parameters are used: Steel: Thermal conductivity 42 W/(m K)
[7] and specific heat capacity 433 kJ/(kg · K) [7]. Aluminium: Thermal conductivity 180 W/(m K) [9]
and specific heat capacity 888 J/(kg · K) [1].
76
Figure 2. Semi finished work piece (left) and formed hybrid bearing bushing (right) [5]
The reference temperature distribution including the temperature jump are well reproduced by the simula-
tions using InTEx elements (Figure 3). For the lower temperature (left in Figure 3) a thermal conductivity
of 0.0105 W/(m K) and for the higher temperatures (center and right in Figure 3) a thermal conductivity
of 0.0075 W/(m K) fit the results from [5] best.† The changing heat conduction is also discussed in [5]
and explained by gap opening between steel and aluminium.
Figure 3. Initial and reference temperature distribution of a hybrid tube according to [5] and simulation results
using InTEx elements for three different initial states
In a different study the thermal behaviour of a friction welded hybrid shaft is examined [6]. This is
assumed to represent the undamaged (D = 0) state as the component is firmly bonded by the joining
process. Some thermocouples are applied close (2.5 mm) to the joining zone and distributed on the
remaining component (TC1 ... TC6 in Figure 4). After induction heating the temperature distribution
is recorded for 10 s. The temperature gradient around the joining zone is very low which indicates a high
thermal conductivity. The global temperature distribution could be reproduced with k = 68 W/(m K),
which represents a perfect serial connection of an equal amount of steel and aluminium (Figure 5):‡
kst kal
kbond = 2 (6)
kst + kal
4 Conclusion
The temperature distribution after a heating process of two different hybrid components was simulated
using InTEx elements for the joining zone. The two components reflect completely firm and fully damaged
joining zone material. In both cases the global temperature distribution could be well reflected; to achieve
this a different thermal conductivity had to be applied which means the thermal conductivity is damage
dependent.
† A joining zone thickness of 10 µm is used here. A different specification of the joining zone thickness (more or less
adjacent material) is accompanied by a different thermal conductivity for the total joining zone material.
‡ The measured temperature gradient could not even be reproduced with a much higher thermal conductivity. Though
the value deduced here is seen as the physical limit and higher values only cause insignificant changes on the temperature
gradient around the joining zone and on the global termperature distribution.
77
Figure 4. Hybrid shaft [6] Figure 5. Initial and reference temperature distribution of a hybrid shaft
according to [6] and simulation results using InTEx elements
Acknowledgments
The results presented in this paper were obtained within the Collaborative Research Centre 1153 “Process
chain to produce hybrid high performance components by Tailored Forming” in the subproject C4. The
authors would like to thank the German Research Foundation (DFG) for the financial and organisational
support of this project.
References
[1] Joachim Ahrendts and Stephan Kabelac. F Technische Thermodynamik. In: HÜTTE-Das Inge-
nieurwissen. Springer, 2012.
[2] B.-A. Behrens, A. Huskic, C. Frischkorn, and D. Duran. Fließpressen hybrider Halbzeuge. Werk-
stattstechnik online 10-2015 (2015), pp. 704–708.
[3] B.-A. Behrens, A. Chugreev, and T. Matthias. Hybride Lagerbuchsen aus Aluminium und Stahl.
Werkstattstechnik online 10-2018 (2018), pp. 691–697.
[4] B.-A. Behrens, B. Breidenstein, D. Duran, S. Herbst, R. Lachmayer, S. Löhnert, T. Matthias, I.
Mozgova, F. Nürnberger, V. Prasanthan, et al. Simulation-Aided Process Chain Design for the Man-
ufacturing of Hybrid Shafts. HTM Journal of Heat Treatment and Materials 74.2 (2019), pp. 115–
135.
[5] Bernd-Arno Behrens, Robert Goldstein, and Anna Chugreeva. Thermomechanical Processing for
Creating Bi-Metal Bearing Bushings. In: Proceedings of the Thermal Processing in Motion. 2018.
[6] Bernd-Arno Behrens, Robert Goldstein, David Guisbert, and Deniz Duran. Thermomechanical
Processing of Friction Welded Steel-Aluminum Billets to Improve Joining Zone Properties (2018).
[7] Deutsche Edelstahlwerke GmbH. Cr-Mn-legierter Einsatzstahl 20MnCr5. 2011.
[8] K. Knothe and H. Wessels. Finite Elemente Eine Einführung für Ingenieure. Vol. 3. Springer, 1999.
[9] thyssenkrupp Materials (UK) Ltd. Material Data Sheet Aluminium Alloy 6082. 2016.
[10] F. Töller, S. Löhnert, and P. Wriggers. Internal Thickness Extrapolation. In: Proceedings in Applied
Mathematics and Mechanics. Vol. 18. 1. Wiley Online Library, 2018.
[11] F. Töller, S. Löhnert, and P. Wriggers. Bulk material models in cohesive zone elements for simulation
of joining zones. Finite Elements in Analysis and Design 164 (2019), pp. 42–54.
[12] F. Töller, S. Löhnert, and P. Wriggers. Modelling Tailored Forming Joining Zones with Internal
Thickness Extrapolation Elements. In: AIP Conference Proceedings. AIP Publishing. 2019.
[13] P. Wriggers. Finite-Elemente-Methode, Skript zur Vorlesung.
78
MS 04: Modern Trends in Numerical Continuation of Stationary Solutions and Application

On Solving Euler-Bernoulli Beams using a B-Spline based

Representation for Random Variables
1
Christoph Eckert ∗
, Michael Beer1 and Pol Spanos2
1 Institute for Risk and Reliability, Leibniz Universität Hannover, Germany
2 L. B. Ryon Endowed Chair in Engineering, Rice University, USA
Abstract. Random quantities of stochastic systems can be approximated by the polynomial chaos repre-
sentation by truncating the series expansion. Ordinarily, the basis of these series are orthogonal Hermite
polynomials, which can be further extended by the Askey scheme of polynomials. In this work, B-spline basis
functions, which are applied successfully in the context of isogeometric analysis, are considered for representing
random variables. As an example, a simple Euler-Bernoulli beam with random beam rigidity is solved within
a Galerkin framework. Further, convergence of the B-spline approach is established by numerical results which
are compared with the multi-element generalized polynomial chaos. As a consequence, the B-spline represen-
tation is a generalization of the Legendre chaos and much more efficient.
Introduction
The usage of polynomial chaos (PC) representation to approximate random processes is widespread in
the area of stochastic mechanics (e.g. [8, 13]). Further, the methodology has a mathematically solid
framework, i.e. for any arbitrary random process with finite second-order moments the original Wiener
polynomial chaos expansion [17] converges in accord with the Cameron-Martin theorem [4]. Further, the
generalization of PC expansions by polynomials of the Askey scheme can be found in [19]. Afterwards,
the generalized PC was further improved by [14, 15, 16] in order to captures the problems of long-
term integration, large perturbation, and stochastic discontinuities. Therefore, the stochastic space is
decomposed in elements and PC approximations are build within each element, which is known as the
multi-element generalized polynomial chaos (ME-gPC).
In 2005, Hughes et al. [10] bridged the gap between computer aided design and engineering by introducing
the methodology of isogeometric analysis (IGA), which successfully has enhanced many deterministic
applications (e.g. [1, 2, 5, 6, 12]). One procedure within the IGA framework is to use non-uniform
rational B-splines, which are prevalent in engineering design processes, as a basis for the analysis. In
addition, it turned out that the nature of B-splines are also very beneficial in terms of analysis procedures.
In this regard, it is remarkable that the notion of IGA is not widely adopted within stochastic frameworks
[3, 9, 11].
In this paper, B-spline basis functions are employed in the truncated PC expansion for approximating
random quantities. Further, an Euler Bernoulli beam with random flexural rigidity is solved by the
B-spline representation using a stochastic Galerkin scheme and faced with the Legendre ME-gPC.
1 B-spline representation for random variables
B-spline basis functions are defined by a knot vector
Ξ = {0, . . . , 0, ξp+2 , . . . , ξn , 1, . . . , 1} = {ξ1 , . . . , ξn+p+1 }, (1)

| {z } | {z }
p+1 p+1
∗ Corresponding author: eckert@irz.uni-hannover.de ORCID iD: 0000-0002-1266-8660
79
consisting of non-decreasing real numbers, where p is the polynomial order and n is the number of basis
functions (e.g. [10]). The ξi ’s are called knots and define element boundaries (knot spans) if they differ.
The B-spline basis functions can then be explained by the Cox-de Boor recursion formula:
(
1 if ξi ≤ u < ξi+1
Bi,0 (u) := for p = 0 , (2)
0 otherwise
u − ξi ξi+p+1 − u
Bi,p (u) := Bi,p−1 (u) + Bi+1,p−1 (u) for p ≥ 1 . (3)
ξi+p − ξi ξi+p+1 − ξi+1
Repeating inner knots in the interior of knot vectors entail in a lower continuity over knot spans, whereas
single knots reveal a C p−1 -continuity. This affects profoundly the number of basis functions and is indeed
the main reason for the effectiveness of the subsequent method.
Next, consider an arbitrary random variable X ∈ L2 (Ω, F, P ) with finite second moment, where Ω is
a sample space, F is an appropriate σ-field on Ω, and P is a probability measure. According to the
polynomial chaos, the B-spline based approximation of X with knot vector Ξ(p, n) is defined by
n
X
e :=
X≈X xi Bi,p (U ), (4)
i=1
where U is uniformly distributed on [0, 1]. The expansion converges, because it suffices that the functions
in use form a basis of the underlying Hilbert space L2 (Ω, F, P) [8, 18]. Therefore, the coefficients xi in
Eq. (4) can be determined by solving a linear algebraic system resulting from the L2 projection:
Ax = b (5)
with
Z Z 1
Ai,j := E (Bi,p (U )Bj,p (U )) = Bi,p (U )Bj,p (U ) dFU = Bi,p (u)Bj,p (u) du (6)
IU 0
Z Z 1
−1 −1 −1
bj := E(FX (U )Bj,p (U )) = FX (U )Bj,p (U ) dFU = FX (u)Bj,p (u) du (7)
IU 0
i, j = 1, . . . , n. The matrix A depends only on the knot vector Ξ of the B-spline basis functions, and can
be stored before the analysis. Further, A is a band matrix with bandwidth p + 1.
2 Euler Bernoulli beam with random stiffness
In this section the versatility and convergence of the aforementioned approach is demonstrated by a
numerical example. Specifically, the ordinary differential equation of an Euler Bernoulli beam with
random stiffness is solved using a stochastic Galerkin scheme and compared to the Legendre ME-gPC by
[14].
15
10
0
0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24 0.26
Figure 1. Solution of Eq. (12) with uniformly distributed beam rigidity W (θ) using four stochastic elements.
Comparison with Monte Carlo simulations (1000 realizations).
80
Figure 2. Comparison of mean errors (left), and variance errors (right) at the beam tip, solved by B-spline rep-
resentation (solid line), and Legendre ME-gPC (dashed line) for uniformly distributed beam rigidity
W (θ).
Consider the Euler Bernoulli beam of length L, clamped at x = 0 and subjected to a deterministic
uniformly distributed load f ≡ 1. The governing equation is given by
d4
EI u(x) = f with u(0) = u0 (0) = 0, (8)
dx4
where the beam rigidity W := EI is assumed to be uniformly distributed in [0.5,1.5], and is specified
by the modulus of elasticity E and the area moment of inertia I. In order to make Eq. (8) numerically
feasible, the spatial and stochastic space of the solution and random input must be discretized. Using
isogeometric subspace leads to
nd
X nd X
X ns
d s d
ũ(x, θ) = ul (θ)Bl,pd
(u(x)) = ui,l Bi,p s
(U (θ))Bl,pd
(u(x)) , and (9)
l=1 l=1 i=1
nw
X
f (θ) =
W w
wk Bk,p (U (θ)), (10)
w
k=1
where u(x) is a linear mapping from the spatial space [0, L] to the parameter space [0, 1]. Applying the
deterministic and stochastic Galerkin procedure to Eq. (8) and using Eq. (9) and (10), yields
Xnd Xns Xnw h i Z d2 d2 d
s s w d
ui,l wk EU Bi,p (U )B j,ps (U )B k,pw (U ) Bl,p (x) B (x) dx (11)
s
L dx
2 d
dx2 m,pd
l=1 i=1 k=1
h iZ j = 1, . . . , ns
s d
= EU Bj,p (U ) f Bm,p (x) dx . (12)
s
L
d m = 1, . . . , nd
Due to the boundary conditions of Eq. (8), the first two coefficients of the vector ul (θ) are equal to
zero. Solving the reduced system, nd · (ns − 2) coefficients of ũ(x, θ) are determined, where as the
control variables ui,nd with i = 1, . . . , ns represents the stochastic beam tip deflection. Since the B-spline
representation is optimal for uniform distributions [7], Eq. (10) is exact for nw = 2, w1 = 0.5, w2 = 1.5,
and knot vector Ξw = {0, 0, 1, 1}. In order to assess the significance of the numerical results obtained from
the B-spline representation the beam problem is threated by a Monte Carlo simulation. The resulting
density functions of the beam tip deflection are plotted in Fig. 1. Comparisons with different B-spline
types show a satisfactory level of accuracy for ps ≥ 2. Further, convergence of mean, and variance of B-
spline representations for the beam tip deflection are shown in Fig. 2 and juxtaposed with multi-element
Legendre chaos. On the one hand, it can be concluded that the B-spline representation generalizes the
multi-element Legendre chaos [7]. On the other hand, the k-refinement property of B-splines (e.g. [10])
results in higher efficiency, e.g. the B-spline representation of order ps = 5 with ten stochastic elements
(ns = 15) results in an increased accuracy for the variance error with less than half of the number of
basis functions in comparison with multi-element Legendre chaos (ns = 31). This effect becomes much
stronger by increasing order.
3 Conclusion
In this paper the potential of a B-spline based representation for random variables has been demonstrated.
Correspondingly, the B-spline approach has been found optimal for uniform input and generalizes the
81
Legendre multi-element polynomial chaos of Wan and Karniadakis [14]. Further, it has been found that
the smoothness property of B-spline basis functions improves significantly the efficiency when decompos-
ing the random space, which even comes to a greater extent if the degree is increased.
As a numerical assertion, a stochastic beam problem has been investigated using Galerkin methods.
Input and output of the system were represented by the new B-spline approach and a better rate of
convergence has been achieved for uniform input in comparison with the multi-element Legendre chaos.
Note that Wan and Karniadakis [14] have stated that the efficiency of ME-gPC is reduced significantly
by the rapidly increasing number of random elements for high-dimensional problems; the results reported
herein indicate that the B-spline representation may overcome this drawback of ME-gPC. The accuracy
of the B-spline technique is also established by pertinent Monte Carlo simulations.
References
[1] F. Auricchio, L. Beirão da Veiga, T.J.R. Hughes, and A. Reali. Isogeometric collocation for elasto-
statics and explicit dynamics. Comput Method Appl M 249-252 (2012), pp. 2–14.
[2] Y. Bazilevs, V. M. Calo, T. J R Hughes, and Y. Zhang. Isogeometric fluid-structure interaction:
Theory, algorithms, and computations. Computational Mechanics 43.1 (2008), pp. 3–37.
[3] G. Bhardwaj, I.V. Singh, and B.K. Mishra. Stochastic fatigue crack growth simulation of interfacial
crack in bi-layered FGMs using XIGA. Comput Method Appl M 284 (2015), pp. 186–229.
[4] R. H. Cameron and W. T. Martin. The Orthogonal Development of Non-Linear Functionals in
Series of Fourier-Hermite Functionals. The Annals of Mathematics 48.2 (1947), p. 385.
[5] J.A. Cottrell, A. Reali, and Y. Bazilevs. Isogeometric analysis of structural vibrations. Comput
Method Appl M 195.41-43 (2006), pp. 5257–5296.
[6] J.A. Cottrell, J.A. Evans, S. Lipton, M.A. Scott, and T.W. Sederberg. Isogeometric analysis using
T-splines. Comput Method Appl M 199.5-8 (2010), pp. 229–263.
[7] C. Eckert, M. Beer, and P. D. Spanos. A B-Spline based Representation for Random Variables
within Stochastic Galerkin Methods (submitted for publication).
[8] R.G. Ghanem and P.D. Spanos. Stochastic Finite Elements: A Spectral Approach. Mineola, New
York: Dover Publications, INC., 2003.
[9] T. D. Hien and H. C. Noh. Stochastic isogeometric analysis of free vibration of functionally graded
plates considering material randomness. Comput Method Appl M 318 (2017), pp. 845–863.
[10] T. J.R. Hughes, J. A. Cottrell, and Y. Bazilevs. Isogeometric analysis: CAD, finite elements,
NURBS, exact geometry and mesh refinement. Comput Method Appl M 194.39-41 (2005), pp. 4135–
4195.
[11] K. Li, W. Gao, D. Wu, C. Song, and T. Chen. Spectral stochastic isogeometric analysis of linear
elasticity. Comput Method Appl M 332 (2018), pp. 157–190.
[12] G. Sangalli, T.J.R. Hughes, L. Beirão da Veiga, F. Auricchio, and A. Reali. Isogeometric Collocation
Methods. Mathematical Models and Methods in Applied Sciences 20.11 (2010), pp. 2075–2107.
[13] G. Stefanou. The stochastic finite element method: Past, present and future. Comput Method Appl
M 198.9-12 (2009), pp. 1031–1051.
[14] X. Wan and G. E. Karniadakis. An adaptive multi-element generalized polynomial chaos method
for stochastic differential equations. Journal of Computational Physics 209.2 (2005), pp. 617–642.
[15] X. Wan and G. E. Karniadakis. Long-term behavior of polynomial chaos in stochastic flow simula-
tions. Comput Method Appl M 195.41-43 (2006), pp. 5582–5596.
[16] X. Wan and G. E. Karniadakis. Multi-Element Generalized Polynomial Chaos for Arbitrary Prob-
ability Measures. SIAM Journal on Scientific Computing 28.3 (2006), pp. 901–928.
[17] N. Wiener. The homogeneous chaos. Amer. J. Math 60.4 (1938), pp. 897–936.
[18] D. Xiu. Numerical Methods for Stochastic Computations: A Spectral Method Approach. Princeton:
Princeton University Press, 2010.
[19] D. Xiu and G. E. Karniadakis. The Wiener–Askey Polynomial Chaos for Stochastic Differential
Equations. SIAM Journal on Scientific Computing 24.2 (2002), pp. 619–644.
82

An Approach to Analyze the Stability of Quasiperiodic

Motions with the Method of Characteristics
1
Robert Fiedler ∗
and Hartmut Hetzler1
1 Engineering Dynamics Group, Institute of Mechanics, University of Kassel, Kassel, Germany
Abstract. The approximation of quasiperiodic motions may efficiently be realized by calculating invariant
manifolds (Tori). Once a stationary solution has been determined, its stability will be of great importance.
This contribution deduces a stability criterion based on the generalized Lyapunov numbers (GLNs), which
can effectively be applied to quasiperiodic motions. Furthermore, the criterion is used to identify the stability
of quasiperiodic motions exhibited by the forced van-der-Pol equation and compared, solely in the stable
region, to an alternative, time simulation based approach.
Introduction
The investigation of quasiperiodic oscillations in engineering applications is a relatively unexplored field

of research, although a well-established theoretical basis is existent in the field of mathematics. At a
first glance, quasiperiodic motions might seem to be a rather artificial case. However, they turn out to
be a widespread phenomenon in engineering which is produces whenever multiple mechanisms arising
from different physical origins excite a dynamical system simultaneously † [2, 5]. The main challenge,
when dealing with quasiperiodic motions, arises from the infinite period length, which results from an
incommensurable frequency basis‡ . Due to that, the use of standard methods (for periodic solutions) is
usually either very disadvantageous or unsuitable, since they often aim on parameterizing the resulting
motion over time. Instead of approaching the problem on one time scale, multiple cyclic time scales
(torus coordinates) are introduced whereat the initial system of ordinary differential equation (dynamical
system in state-space) transfers to a system of partial differential equations (invariance equations). The
solution of the invariance equation is an invariant manifold (object in the state-space), which is filled
densely by a quasiperiodic motion. Once calculated, the invariant manifold represents an unperturbed,
stationary motion of a dynamical system. For any stationary solution informations about the stability
are of great importance
The definition of orbital stability for quasiperiodic motions in the sense of Lyapunov is theoretical
possible [3] but almost always not applicable when dealing with dynamical systems which cannot be
solved analytically. In addition, algebraic stability criteria derived from linear maps are in the context
of quasiperiodic motions not available§ . This problem stems from the infinite periodic length possessed
by a quasiperiodic solution in time domain. Since stability criteria are often derived from the linearized
equation of motion, the exponential grow and/or decay rate result in singularities and/or infinitely large
values. Nevertheless, there are methods existent which aim on approximating the Lyapunov exponents
of arbitrary solutions [1]. The drawbacks, exhibited by these methods, are on the one hand the high cal-
culation effort and on the other hand that they often base on time integrations, whereby an identification
of instabilities is not possible. This contribution aims on deriving a general and time efficient approach
to identify the stability of quasiperiodic motions.
∗ Corresponding author: robert.fiedler@uni-kassel.de

† Exceptions may be present when synchronization occurs.
‡ Two incommensurable (rational independent) frequencies are for example ω = 1 and ω = π.
1 2
§ In the context of equilibrium points or periodic solutions, the stability is usually identified by an eigenvalue analysis
of the Jacobian or the Monodromy matrix, respectively.
83
The theory of quasiperiodicity is extensive and a detailed reproduction would exceed the scope of this con-
tribution. Interested readers are e.g. referred to [3] for a comprehensive explanation of the mathematical
theory. When dealing with periodic motions, a boundary value problem can be formulated to identify a
solution. The main problem, when dealing with quasiperiodic motions, arises from the parameterization
over time. Due to the infinite period length,
h the resulting
i boundary value problem has to be solved on
an infinite, one dimensional interval t ∈ 0, lim T . Instead of directly using the time, a parametriza-
T →∞
tion over different time scales (θ1 = ω1 t, . . . , θp = ωp t) is usually used, whereby the different time scales
are connected to the incommensurable frequency basis (ωi , i = 1, . . . p). The resulting parametrization
θ = [θ1 , . . . , θp ] consists of cyclic coordinates (torus coordinates), whereby a parametrization over a finite,
p-dimensional interval is achieved. For example, consider an dynamical system in state-space
ż = f (z, Ωt, γ), z(t) ∈ Rn , Ω ∈ Rm , γ ∈ R, (1)
where z is the state-space vector, Ω is a vector of incommensurable forcing frequencies and γ is the
bifurcation parameter. Parameterizing Eq.(1) on a torus instead over time is a change of the independent
variable from t to θ ¶ : z(Ω1 t, . . . Ωm t, ωm+1 t, . . . , ωp t)k = Z(θ1 , . . . , θp ). With this one can derive the
invariance equation
Xm Xp
∂Z ∂Z
Ωi + ωi = f (Z, θ̃, γ), θ̃ = [θ1 , . . . , θm ] (2)
i=1
∂θi i=m+1
∂θi
where besides the torus function Z, the autonomous frequencies ω = [ωm+1 , ..., ωp ] are unknown. Analo-
gous to periodic solutions, the necessity of a phase condition is given, whereby depending on the solution
strategy different approaches are possible (e.g. [4]).
1.1 Application of the method of characteristics
Since the focus of this contribution lies on the stability assessment, it is assumed that a quasiperiodic
motion zq (t) (Zq (θ(t)), respectively) is known. Perturbing the quasiperiodic solution (z = zq + ∆z) and
performing subsequently a Taylor-expansion, the linearized equations of motion in time domain are
obtained
2 ∂f
∆ż = j(t)∆z + O(∆z ), j(t) = , j(t) ∈ Rn×n , (3)
∂z zq
where j is the Jacobian in time domain: since j depends on the quasiperiodic motion it is quasiperiodic,
too. Converting the parametrization from time to torus domain ([t, zq , j(t)] → [θ, Zq , J(θ)]) and neglecting
the higher order terms, yields a linear system of partial differential equations
m
X Xp
∂∆Z ∂∆Z
Ωi + ωi = J(θ)∆Z. (4)
i=1
∂θi i=m+1
∂θi
One should note, that at this point the frequencies ωi , i = m + 1, ..., p are known∗∗ . The system of
linear partial differential equations (4) can be solved by using the method of characteristics, wherewith
a periodic boundary value problem is converted into an initial value problem. The resulting system of
ordinary differential equations reads:
dθi
= Ωi , i = 1, . . . , m
dt d∆z
= J(θ(t))∆z, ∆z(0, s) = ∆z0 , (5)
dθi dt
= ωi , i = m + 1, . . . , p
dt
where ∆z0 are initial values and the parametrization changes from θ to [t, s], t ∈ R, s ∈ Rp−1 . The
variable t corresponds to the time of the underlying dynamical system and s parametrizes the initial curve.
Using this formulation, the evolution of a perturbed solution over time ∆z(t) can be combined with the
torus formulation J(θ), whereby the position on the torus can always be identified by the evolution of
the torus coordinates θ(t). Through the definition of an initial curve a unique solution can be obtained.
¶ I.e.z(t) = z(ω1 t) = Z(θ1 ).
k Note that the first m arbitrary frequencies ωi , i = 1, ..., m are substituted by the known incommensurable forcing
frequencies Ω.
∗∗ Through the calculation of the invariant manifold by Eq.(2) and an additional phase condition (e.g [4]).
84
∆zi (t, 0) = ∆zi (t, send ) Note that the choice of the initial curve is not
~e∆zi
2π completely arbitrary since it has to satisfy the
∆zi (0, s) ~eθ2
~eθ1 c
2π transversality condition. Defining a set of linear
ct eristi independent vectors Φ = [∆z1 | . . . | ∆zn ] with
ra ∆z i end , s)
(t
Cha the identity matrix as initial value Φ(0, s) = I, a
~es Zq,i (θ1 , θ2 ) normalized fundamental system is defined. Since
~et
Eq.(2) implies parallel dynamics†† and the fre-
Figure 1. Schematic depiction of the approach with two quency basis [Ω, ω] is known, the time interval
torus coordinates (θ1 , θ2 ). Note that only one [0, tend ] on which the characteristic system has to
perturbed component (∆zi ) is depicted. be solved can be determined. Solving the charac-
teristics on the defined time interval leads, due to
the underlying quasiperiodic problem, to a linear map which can map arbitrary perturbations over a cyclic
boundary (cf. figure 1). Choosing the initial curve orthogonal to θ1 and solve Eq.(5) for a normalized
fundamental system, a map
∗
∗

∆Z(2π, θ̃ ) = Φ [2π, 0], [θ̃ , θ̃] ∆Z(0, θ̃), θ̃ = [θ2 , ..., θp ] (6)
is found, which describes the behavior of an arbitrary perturbation over the cyclic boundary. Here,
θi∗ = θi + 2π ωω1i . It is important to note, that an eigenvalue analysis of Φ is not applicable, since a point
on the initial curve ∆Z(0, s) is not mapped onto itself‡‡
1.2 Stability criterion for quasiperiodic motions
Following, the theory of generalized Lyapunov numbers is used to deduce an applicable criterion to
identify the stability of quasiperiodic motions. For a detailed view of the theory see [6, p. 56 – 69]. Since
an invariant manifold possesses the particular property that orbits never leave it, solely the dynamics
normal to the invariant manifold are of interest. Therefore, with the standard (Euclidean) norm, the
linearized space is split
T Rn |M = T M ⊕ N (7)
into a tangent (T M) and normal bundle (N ). Through this splitting, the following linear operators can
be constructed for the linearized flow
A(t) ≡ φ(t, t0 ) : T∆z(t0 ) M → T∆z(t) M
(8)
B(t) ≡ q(t)φ(t, t0 ) : N∆z(t0 ) → N∆z(t) ,
where q(t) : T Rn |M → N indicates an orthonormal projector of a vector onto the normal bundle and φ is
the fundamental solution. Both defined in time domain. By means of the linear operators, the following
holds:
v(t) = A(t)v(t0 ) v ∈ TM
(9)
w(t) = B(t)w(t0 ) w ∈ N.
Considering the linear operator of the normal space B(t), the first generalized Lyapunov number (GLN1)
can be defined
1 ||B(t)w(t0 )||
ν = lim ||B(t)|| t , with ||B(t)|| ≡ sup . (10)
t→∞ w(t0 )∈N ||w(t0 )||
Through Eq.(6) the fundamental solution is known over a cyclic boundary. Transforming the expression
of the GLN1 into torus domain ([t, q(t), φ(t)] → [θ, Q(θ), Φ(θ)]) results in
N N t1 (
Y end < 1 stable

ν = lim Qi Φi → (11)
N →∞
i=1
> 1 unstable
whereby an applicable and time efficient criterion for the stability of quasiperiodic motions is formulated.
Note that the GLN1§§ is independent of the splitting Eq.(7), constant on orbits and converges over time.
†† Note that the torus coordinates are linear functions of time θi = ωi t.
‡‡ The ωi
offsets (2π ω ) have to be considered.
1
§§ And the second generalized Lyapunov number as well.
85
2 Application to the forced van-der-Pol equation
In order to show the practicability of the proposed approach, it is applied to quasiperiodic solutions of
the forced van-der-Pol equation

x00 + ε x2 − 1 x0 + x = fˆ cos(Ωτ ), (12)
where (.)0 is the derivation with respect to the dimensionless time τ , ε controls the nonlinearity and
fˆ and Ω are the forcing amplitude and dimensionless frequency, respectively. Following, the excitation
parameters are fixed (fˆ = 0.6, Ω = 1.5) and the bifurcation parameter is the nonlinearity (ε ∈ [−1, 1]).
The quasiperiodic solutions are calculated by means of a finite difference method (FDM). The resultant
quasiperiodic solution hast two base frequencies, one known non-autonomous (Ω) and one unknown
autonomous (ω). In order to compare the identified stability behavior quantitatively, the Lyapunov
exponents (LEs) are calculated with a time simulation based method [1].
1.5
σG = ln(ν)
1 σ1
σ2
0.5 σ3
0
σ
−0.5
−1
−1.5
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
ε
Figure 2. Comparison of Lyapunov exponents: Continuation (σG = ln(ν)) and time simulation (σ1 , σ2 , σ3 ).
The investigation of Eq.(12) depicted in Fig.2. confirms the expected behavior. Since the autonomous
state-space (z ∈ R3 ) is spitted into T M ∈ R2 and N ∈ R the normal bundle is one dimensional,
whereby one LE represents its contraction/expansion rate (σ3 ). Furthermore, two LEs are equal to zero,
which identifies tangential directions (σ1 , σ2 ). The results show, the time simulation based method can
approximate the LEs until ε = 0. At this point the stability property changes and solely the proposed
method can identify the instability of the quasiperiodic motion. Hence, the proposed method can not
only predict the general stability property of invariant manifolds, in addition it does the prediction more
efficient with reference to time and computational effort.
References
[1] F. Christiansen and H. H. Rugh. Computing Lyapunov spectra with continuous Gram - Schmidt
orthonormalization. Nonlinearity 10.5 (1997), p. 1063. issn: 0951-7715.
[2] D. Guo, F. Chu, and D. Chen. The Unbalanced Magnetic Pull and its Effects on Vibration in a Three-
Phase Generator with Eccentric Rotor. Journal of Sound and Vibration 254.2 (2002), pp. 297–312.
issn: 0022460X.
[3] A. M. Samoilenko. Elements of the Mathematical Theory of Multi-Frequency Oscillations. Vol. 71.
Mathematics and Its Applications, Soviet Series. Dordrecht: Springer, 1991. isbn: 9789401055574.
doi: 10.1007/978-94-011-3520-7.
[4] Frank Schilder, Hinke M. Osinga, and Werner Vogt. Continuation of Quasi-periodic Invariant Tori.
SIAM Journal on Applied Dynamical Systems 4.3 (2005), pp. 459–488. doi: 10.1137/040611240.
[5] Sunil K. Sinha. Rotordynamic analysis of asymmetric turbofan rotor due to fan blade-loss event
with contact-impact rub loads. Journal of Sound and Vibration 332.9 (2013), pp. 2253–2283. issn:
0022460X.
[6] Stephen Wiggins. Normally Hyperbolic Invariant Manifolds in Dynamical Systems. 105 (1994).
86

Initialization of the continuation of stick-slip vibrations

for a two dof friction oscillator
1
Jonas Kappauf ∗
and Hartmut Hetzler1
1 Engineering Dynamics Group, Institute of Mechanics, University of Kassel
Abstract. Within this contribution, approximation schemes for calculating periodic solutions for nonlinear
self-excited systems are studied. For applying the Harmonic Balance method numerically, suitable initial
conditions when applying Newton-like schemes are required. Hence, a procedure based on homotopy methods
is presented for initializing the numerical calculation. In particular, a fixed-point homotopy map is constructed.
This procedure is applied to calculate self-excited periodic stick-slip vibrations of a two degree of freedom
friction oscillator. To this end, the influence of the harmonic order H on the homotopy path is studied and
an outlook is given.
Motivation
Focussing on mechanical systems with strongly nonlinear or non-smooth forces, the initialization of nu-
merical approximation schemes, e.g. for calculating periodic solutions, is the crucial point. As a method
for approximating periodic limit cycles, that provides the ability dealing with non-smoothness in the
equations of motion, the Harmonic Balance Method (HBM) is adressed.
For solving the resulting nonlinear algebraic HBM equation system numerically, an appropriate initial
guess for the Newton-Raphson scheme is appreciated to achieve good convergence behaviour. Focussing
on self-excited vibrations, usually less information may be taken from the linear system, hence the base
frequency of the limit cycle is unknown. However, suitable methods, e.g. homotopy methods offering the
ability to transform a simple or simplified version of an algebraic system, where the solution is known, to
the original one. That results in a expansion of the basin of a attraction for a local solution of the HBM
equations. In particular, the applicability of a fixed-point homotopy method approach is addressed to
calculate periodic limit cycles of a self-excited two degree of freedom (dof) friction oscillator with velocity
depending friction characteristic showing stick-slip vibrations.
1 Basic equations of the Harmonic Balance
For calculating periodic limit cycles, the Harmonic Balance method is a generic tool assuming the solution
of a dynamical system can be expressed as a Fourier series of the form
X
H
X0 jkωt
x(t) = + Re Xke . (1)
2
k=1
Substituting τ = ωt and inserting eq. (1) into the equations of motion, denoted by f eom (ẍ, ẋ, x) = 0, a
Galerkin projection plus a scalar equation pc leads to the set of algebraic equations
Z
1 2π
Rk = f eom (X i , ω, τ )e jkτ dτ, i, k = 0, 1, ..., H (2a)
π 0
pc = g(X i , ω) (2b)
∗ Corresponding author: kappauf@uni-kassel.de ORCID iD: 0000-0001-5142-4530
87
for calculating the coefficients X i and the base frequency ω of eq. (1). Here, Rk denotes the coefficients
of the residual related to the harmonic order k while pc is the scalar phase condition, see [1].
However, the calculation of periodic limit cycles require adequate initial conditions when applying eqs. (2)
to a classical Newton-Raphson scheme.
2 Homotopy methods
As mentioned above, generating initial conditions for calculating solutions in an straight-forward way, in
this contribution homotopy methods are applied. Besides expanding the basin of attraction for solving
the algebraic equations [2], homotopy methods are able to detect isolated solution branches. In both
cases, a homotopy map H λ is defined, where
H 0 (x; µ) = G(x; µ) (cheap problem)

H λ (x; µ) = 0, for 0 < λ < 1 and (3)
H 1 (x; µ) = F (x; µ) (expensive problem)
holds. For embedding the nonlinear algebraic equation systems, here eq. (2), of the form
F (x; µ) = 0, x ∈ RN , µ∈R (4)
into the framework of eq. (3), a convex approach is selected, where
H λ (x; µ) = λF (x; µ) + (1 − λ)G(x; µ) (5)
represents a linear superposition related to λ for fixed µ and H λ = 0 is assumed. Suppose F , G are
smooth functions, this map implicitly defines an smooth solution path γ = (x> , λ)> , that can be traced
using path classical continuation methods, e.g. predictor-corrector schemes [1]. To ensure that γ is
reaching a solution of eq. (4), both γ has finite length and the implicit function theorem holds is assumed
[4]. Moreover a careful selection of G(x; µ) increases efficiency of the method.
Within this contribution, the fixed-point homotopy is applied [3, 4], where G(x; µ) has only one solution
x = a and is represented by
G(x; µ) = (x − a), a ∈ RN . (6)
1.2
3 Application
friction force fµ [-]
Following this procedure, periodic limit cycles of a 2 dof oscil- x2

lator with friction law showing in Fig. 1 are calculated. As an 0.6 x1 m2
initial value for the homotopy map H λ , the solution a for an d
k1
uncoupled linear system with two degrees of freedom showing
purely harmonic oscillations with ω = 1 at λ = 0 is given. The m1
k2
periodic limit cycle of the system, shown in Fig. 1 is calculated µ(vrel )
vb
for variable harmonic order H ∈ {1, 2, 3, 4} at constant belt
speed vb = 0.35. 0 0.4 0.8
relative velocity vrel
The homotopy maps are obtained via classical predictor-
corrector scheme wit additional step size control taken from [1] Figure 1. Friction oscillator with two de-
and pictured in Fig. 22b. For increasing number of unknowns, grees of freedom incl. velocity
the homotopy path may include more turning points. Moreover, dependent friction coefficient
certain areas occur, where small variation of λ leads in a large
change of solution components, that results in higher computa-
tional effort. Especially H = 4 leads to folds in the bifurcation
path near λ ≈ 0.78. Zooming into that domain offers that there
are more iterations of the continuation algorithm required.
88
0.5 1.1
x1 H=1 H=2 x1
RMS of periodic solution x [-]

0.4 x2 1 x2
0.3 vb 0.9 H=3
0.2 0.8
velocity x0i [-]
H=4
0.1 0.7
0 H=3 H=1
0.6
−0.1
0.5
−0.2
−0.3 H=4 0.4
H=2
−0.4 0.3
−0.5 0.2
−3 −2 −1 0 0 0.2 0.4 0.6 0.8 1
deflection xi [-] homotopy parameter λ [-]
(2a) Approximated stick-slip limit cycle of the system (2b) Homotopy maps of the global fixed-point homotopy
m2 for different values of H.
shown in Fig. 1, where k1 = 1, k2 = 1.5, = 0.5
m1
and d = 0.02.
Nevertheless, the solution of the complex problem, namely the stick-slip limit cycle of the mechanical sys-
tem, is calculated at λ = 1, however the requirements on the continuation procedure are quite challenging,
especially relating to efficiency and computational effort.
4 Conclusion
This contribution aims to establish a straight forward procedure for expanding the basin of attraction
for the Harmonic Balance method when approximating periodic limit cycles of self-excited systems. A
possible convex design of a homotopy map was given and the implicit defined curve γ was traced via
classical path tracking techniques. In the particular case of a two dimensional friction oscillator, solutions
at a certain parameter value of the belt speed vb were obtained. It was pointed out that for increasing
number of unknowns in the nonlinear equation system, the complexity as well as the computational effort
of the solution paths increase as well. In a current state of research, it is of interest to study and apply
some criteria, that are already used in chemical engineering [3], to affect the shape of the homotopy path
γ targeting the solution of the complex problem.
References
[1] W. Vogt B. Marx. Dynamische Systeme: Theorie und Numerik. Springer Verlag, 2010.
[2] J. Kappauf and H. Hetzler. Bifurcations and limit cycles due to self-excitation in nonlinear systems
with joint friction: Initialization of isolated solution branches via homotopy methods. Proceedings in
Applied Mathematics & Mechanics (2019).
[3] M. Kuno and J.D. Seader. Computing all real solutions of Systems of nonlinear equations with a
Global Fixed-Point Homotopy. I. & e. c. r. (1988).
[4] L.T. Watson. A survey of probability-one homotopy methods for engineering optimization. Depart-
ment of Computer Science, Virginia Polytechnic Institute & State University (1990)).
89
90
MS 05: Multi-field problems in porous media mechanics

Thermo-hydro-mechanical modeling of liquid freezing

in saturated porous media
1
Abdel Hassan Sweidan * , Yousef Heider2 and Bernd Markert1
1 Institute of General Mechanics, RWTH Aachen University, Aachen, Germany
2 Department of Civil Engineering and Engineering Mechanics, Columbia University, New York, USA
Abstract. In the underlying research work, a thermo-hydro-mechanical (THM) finite element model is introduced to investi-
gate the freezing-thawing process in a non-isothermal fluid-saturated porous medium. In this, a bi-phase macroscopic model
considering a granular porous matrix and a fluid phase is proposed to account for the temperature development, fluid-phase
transitions and volume deformations due to the ice formation based on a macroscopic phase-field-porous media approach. The
theory of porous media (TPM) provides a well-founded continuum mechanical basis for the description of deformable, fluid-
saturated non-isothermal porous solid aggregates. In particular, the ice and water constituents are treated as a single pore fluid
and described by a unified formulation, where the phase-field method (PFM) is employed for the description of the phase tran-
sition between both components. The PFM is a diffuse interface model that relies on the specification of the free energy density
function, which is the main driving force for phase transition and employs a phase-field variable to describe whether the material
is ice or water. Significant virtue of the diffusive phase-field approach is that it is easy to implement and eliminates numerical
challenges associated with modeling of moving sharp boundaries. Finally, a numerical example is presented to demonstrate
the ability and usefulness of the model in describing the freezing process in a saturated porous solid under thermal loading.
Introduction
Freezing and thawing processes in fluid-saturated porous media are of great interest in many branches of engi-
neering. Porous materials, like road pavements and concrete, are usually subjected to frost damage in cold regions
where periodic freezing occurs. Moreover, artificial ground freezing techniques are widely used for ground-water
control and temporary excavation support during tunnel construction in non-cohesive soils. Ice formation in porous
media is characterized by a coupled heat and mass transport and is accompanied by ice expansion (9% volume di-
lation referred to the volume of liquid) and considerable deformations [1].
In order to capture such multi-physical phenomena on an engineering scale, a thermo-hydro-mechanical model
consisting of a porous solid matrix and a pore-fluid phase representing solid ice and liquid water is proposed. The
model is based on a coupled phase-field-porous media approach and accounts for the phase change, volume expan-
sion and solid deformation during freezing. In this, the macroscopic theory of porous media (TPM) [6] is used for
the description of the deformable, heterogeneous porous solid with the phase changing fluid constituent described
by a unified formulation employing a phase-field model (PFM) [2]. The development of a thermodynamically
consistent description of phase-change processes in porous media based on the necessity of satisfying the require-
ments of the entropy inequality of the TPM has been presented in [1]. However, the phase change process in this
article is described by the mass transfer from one phase to another depending on the mass production term.
The treatment of phase change processes in a continuum description by introducing a singular front (interface)
has already been tackled by many researchers [8]. Most popular interface capturing methods are the level-set,
volume-of-fluid, and phase-field method (PFM). In this regard, the energy-based PFM [11] introduces a phase-
field variable φF to indicate the phases of the material, e.g. φF “ 0 for the solid and φF “ 1 for the liquid
states. The PFM allows for a smooth variation between these distinct values across the interface, resulting in a
finite interface thickness. Such diffusive interface treatment of the freezing front can easily be implemented in
finite-element codes as no explicit front tracking or application of boundary conditions at the interface is required.
Therefore, we employ the PFM as an alternative method for the traditional mass transfer method that adopts the
mass production term for phase change in TPM. The study combines the concepts of the TPM and the PFM for
the description of the freezing process in saturated porous media, where the thermodynamic consistent constitutive
* Corresponding author: sweidan@iam.rwth-aachen.de
91
equations of the model together with the evolution equation of the phase-field variable are derived by exploiting
the entropy inequality. Finally, a 2D numerical example is presented to demonstrate the ability of the proposed
model in describing the freezing process including volume expansion in fluid-saturated porous media.
In the framework of the TPM [6], a continuum-mechanical description of the material ϕ consisting of incompress-
ible and immiscible constituents ϕα (α “ S : solid, α “ F : fluid) is being presented. Here, the fluid occupying
the pore space is the water in its ice (α “ I) or liquid (α “ L) phases. Thus, the volume fractions, the saturation
condition and the density dependencies, respectively, read:
nα “ dv α {dv , n S ` nF “ 1 , ρα “ nα ραR . (1)
Here, dv α and dv are the partial and total volume elements, and ραR and ρα are the partial and material densities,
respectively. The solid matrix is described in a Lagrangian manner via the solid displacement uS and velocity vS ,
while the fluid is described by a modified Eulerian setting via the seepage velocity vector wF “ vF ´ vS .
To obtain thermodynamically admissible constitutive equations for the stresses and the production terms, the en-
tropy inequality should be fulfilled. Considering no mass transfer between the constituents, the time derivative of
the saturation condition (Eq. (1)2 ) with respect to the solid deformation is multiplied by the Lagrange multiplier Λ
and added to the Clausis-Duhem in equality
TF
TS hkkkkkkkkikkkkkkkkj
E
p̂F
hkkkkkikkkkkj
E
S F
hkkkkkkkkikkkkkkkkj
E
θ θ
rTS ` ΛnS Is ¨ LS ` F rTF ` S ΛnF Is ¨ LF ´ rp̂F ´ Λgrad nF s ¨ wF
θ θ
3 α S
θ S
θ F 3 αF θS (2)
´ρS pθS q1S rη S ` Λ SR s ´ F ρF pθF q1F rη F ` Λ S F R s ´ F ρF pψ F q1F ´ ρS pψ S q1S
ρ θ θ ρ θ
S
1 S θ θS
´ S q ¨ grad θS ´ F 2 qF ¨ grad θF ` ε̂F p F ´ 1q ě 0 ,
θ pθ q θ
where η α is the constituent mass-specific entropy and θα is the absolute Kelvin temperature of ϕα . Λ could be
identified as Λ “ θS {θF p for the case of non-isothermal constituents with p being the pore pressure. Tα “ pTα qT
represents the symmetric partial stress tensor and Lα “ grad vα is the spatail velocity gradient. εα is the internal
energy, qα is the heat influx and ε̂α is the direct energy production representing the local heat exchange between
ϕα . p̂α is the volume-specific local interaction force between ϕF and ϕS , where p̂S ` p̂F “ 0 must hold.
Employing the principle of phase separation [4], where the free energy of an individual component of the overall
aggregate does only depend on its own constitutive variables as on the microscale, the Helmholtz free energy of
the linear thermoelastic solid skeleton is written as
ˆ S
˙
1 S S θ
S S S S 2 S S S S
ρ0S ψ pθ , εS q “ µ pεS ¨ εS q ` λ pεS ¨ Iq ` m ∆θ pεS ¨ Iq ´ ρ0S Cp θ ln S ´ ∆θ S
. (3)
2 θ0
Following basic papers on phase-field models for a single-phase material [10] and the arguments provided by [1,
12], the Helmholtz free energy for the fluid mixture ϕF including the phase change is proposed as
1 2 1
ρF R F F F F
0F ψ pθ , φ , grad φ , εF q “ |grad φ|2 ` W gpφF q ` µF pεF ¨ εF q ` λF pεF ¨ Iq2
2 ˆ ˙ 2
θ F ` ˘
`mF ∆θF pεF ¨ Iq ´ ρF R F
0F Cp θF ln F ´ ∆θF ´ 3 1 ´ ppφF q k I αpF pεF ¨ Iq (4)
θ0
LF
´ppφF qρF
0F
R
F
F
∆θM ´ ρF R I F
0F η0I ∆θM .
θM
Herein, µα and λα are the Lamé constants, κα “ 2{3 µα ` λα is the compression modulus, mα “ ´3 κα αF is
the stress–temperature modulus and αpF is the volumetric expansion coefficient due to phase change of ϕF . Cpα
represents the specific heat capacity. LF and η0I
I
are the latent heat of fusion and the reference value of the specific
entropy of ice at the freezing point, respectively. ∆θα and ∆θM α
are the temperature variations compared to the
reference and melting temperatures θ0 and θM , respectively. The subscript {0 α} refers to quantity at time t0.
F F
, W are the gradient energy coefficient and energy hump parameter, respectively. The evaluation of the entropy
92
inequality is carried out according to the procedure of Coleman & Noll [3]. By doing so, the constitutive relations
and the phase-field evolution equation are chosen as
µF R F 2
p̂F
E “´ pn q wF , qS “ ´HS grad θS , qF “ ´HF grad θF ,
ε̂F “ kθVFS S
pθ ´ θF q ,
KS
” ρF R LF ı (5)
pφF q1F “ M 0FF ∆θM F 1 F
p pφ q ´ W g 1 pφF q ` 2 ∇2 φF ,
θM
where µF R ą 0 is the dynamic fluid viscosity and K S ą 0 represents the isotropic` intrinsic permeability ˘of ϕS .
gpφF q “ pφF q2 p1 ´ φF q2 is a double well energy function and ppφF q “ pφF q3 6 pφF q2 ´ 15 φF ` 10 is an
interpolating function with g 1 pφF q and p1 pφF q denoting their derivatives with respect to φF , respectively. The
interface mobility M together with W and could be found by relating them to the material parameters and the
interface width as in [2]. Moreover, to consider a change in the permeability K S due to the deformation of ϕS ,
the relation provided by Markert [9] is chosen. In the sense of Fourier’s law of heat conduction, Hα “ nα H αR I
represents the partial heat conductivity tensor. kθVFS
“ 3 nF {rF kθSFS
is a positive volume-specific heat transfer
coefficient with kθS being the heat transfer coefficient at the interface between ϕS and ϕF , rF is the average pore
FS
channel radius and rS is the characteristic solid particle radius.

Under small-strain conditions and based on Eq. (3), the solid effective stress reads
ρS
TSE “ r2 µS εD S S S
S ` κ pεS ¨ IqI ` m ∆θ Is , (6)
ρS0S
where εDS is the deviatoric strain tensor. The dependencies of ρ and n on the solid deformation and thermal
S S
expansion can be found following the procedure provided by [5] and [7]. Based on Eq. (4), the effective fluid stress
reads
ρF ” F ı ρF ” F L ` ˘ Iı
TF F F D F F F
E “ F R m I∆θ ` 2 µ εF ` κ IpεF ¨ Iq ´ 3 k αp I “ F R φ T E ` 1 ´ φ
F
TE . (7)
ρ0F ρ0F
In this, φF is employed to allow for a smooth variation of an arbitrary thermo-mechanical
` ˘property % between the
liquid and frozen phases of ϕF during the phase transition, e.i. %F “ φF%L ` 1 ´ φF %I . Based on (7), the
linear description of ϕF in its liquid state (φF “ 1) follows by setting µL “ 0. Moreover, the incompressible
liquid case is obtained by letting pεL ¨ Iq Ñ 0. The case of incompressible Newtonian fluid is simply recovered by
letting T L
E “ 2n µ
F FR
DF . For each computational cell, the displacement is obtained by integrating the velocity
over time, i.e. puF q1S “ vF .
The governing partial differential equations (PDEs) with primary unknowns {uS , vF , p, θF , θS , φF } are summa-
rized as follows:
Mass balance for ϕF : pρF q1F ` ρF div vF “ 0 ,
pnF q2 µF R
Momentum balance for ϕF : 0 “ div TF
E ´ n grad p ` ρ g ´
F F
wF ,
KS
“ S θS ‰
Mixture momentum balance : 0 “ div tTSE ` TF E ´ n p F ´ 1q ` 1 p Iu ` rρS ` ρF s g ,
θ (8)
Energy balance for ϕF : ρF CpF pθF q1F ` ρF LF p1 pφF qpφF q1F “ TF ¨ LF ´ div qF ` ε̂F ,
Energy balance for ϕS : ρS CpS pθS q1S “ TS ¨ LS ´ div qS ´ ε̂F ´ p̂F

E ¨ wF ,
” ρF R LF ı
Phase-field equation : pφF q1F “ M 0FF ∆θM F 1 F
p pφ q ´ W g 1 pφF q ` 2 ∇2 φF .
θM
In order to avoid lengthy formulation of the balance equations, we provided simplified forms through neglecting
the terms that has a minor effect on the solution. This includes omitting the terms related to the thermal expansion,
pressure and elastic strains in Eq. (8)4 . Moreover, the contribution of elastic strains was neglected in the phase-
field evolution (Eq. (8)6 ), as the primary purpose of φF is to track the interface based on the thermal driving force.
To overcome stability challenges associated with the assumption of materially-incompressible fluid, a pseudo-
compressibility-based scheme via Laplace of the pressure is applied in this work. When dealing with special
cases where no solid deformation is considered, KS in Eq. (8)2 could be replaced by a modified version KφS “
p1 ´ φF qKminS
` φF K S , where Kmin S
! K S is a very small positive constant. This relation is introduced to ensure
that the fluid velocities reduce gradually from finite values (for KφS “ K S ) in the liquid state of ϕF to almost zero
values (for KφS “ Kmin S
) in the frozen state.
93
2 Numerical results and discussion
In this section, a 2D numerical simulation is performed using the Finite Element Method (FEM) in order to present
the application of the proposed model. The geometry and boundary conditions of the sample are shown in Fig.
1. Specifically, the model is applied to a solid specimen (cement stone) having a width L “ 0.15 m and a height
H “ 0.075 m. All the unstressed surfaces are adiabatic to the environment. The system is initially at a temperature
θ0α “ 293 K with a solid volume fraction nS0S “ 0.5. The bottom of the specimen is cooled in-line from 293 K to
253 K within 4 hours and then is held constant within 1 hour. All the four walls of the specimen are impermeable
and no-slip boundary conditions are imposed (v̄F “ 0). The thermo-physical properties of the porous material,
ice and water are adopted from [1].
Bθ̄ α
“ 0
v̄F “ 0 By
˝C φF
111111111111111111111111111111111111111
000000000000000000000000000000000000000
000000000000000000000000000000000000000
111111111111111111111111111111111111111 1 1
000000000000000000000000000000000000000
111111111111111111111111111111111111111
000000000000000000000000000000000000000
111111111111111111111111111111111111111
α L
000000000000000000000000000000000000000
111111111111111111111111111111111111111 α
Bθ̄ 000000000000000000000000000000000000000 Bθ̄
111111111111111111111111111111111111111
“ 0 111111111111111111111111111111111111111
000000000000000000000000000000000000000 “ 0
Bx
000000000000000000000000000000000000000
111111111111111111111111111111111111111
000000000000000000000000000000000000000
111111111111111111111111111111111111111 Bx
000000000000000000000000000000000000000
111111111111111111111111111111111111111
000000000000000000000000000000000000000
111111111111111111111111111111111111111
H ´10 0.5
000000000000000000000000000000000000000
111111111111111111111111111111111111111
y
v̄F “ 0 111111111111111111111111111111111111111
000000000000000000000000000000000000000 v̄ “ 0
000000000000000000000000000000000000000 F
111111111111111111111111111111111111111
000000000000000000000000000000000000000
111111111111111111111111111111111111111
000000000000000000000000000000000000000
111111111111111111111111111111111111111
000000000000000000000000000000000000000
111111111111111111111111111111111111111
θ̄ α “ θc
000000000000000000000000000000000000000
111111111111111111111111111111111111111
x
000000000000000000000000000000000000000
111111111111111111111111111111111111111 ´20 0
y (a) (b)
Figure 1. 2D geometry and bound- Figure 2. (a) Temperature and (b) phase-field evolution con-
ary conditions tours of the deformed porous solid at time t = 5 h
Fig. 2 shows the liquid/ice interface position and the accompanied volume deformation after 5 hours. The volume
increase is mainly subject to the moving freezing front inside the porous solid, however a thermal dilatation is
observed during the cooling phase in-between 293-273 K. At θF ď 273 K, the phase change from liquid to ice
occurs and a significant volume deformation of the specimen due to ice freezing is observed. These results agree
with the observations in [1]; however, in their work, the isothermal equilibrium model was used and the phase
change process was described by the mass transfer via the mass production term. Moreover, the volume expansion
due to phase change appeared to be only in the lateral direction.
3 Conclusion
In this study, a bi-phasic macroscopic model has been presented for the simulation of freezing process in liquid-
saturated porous media. The model is based on a coupled phase-field-porous media approach and allows to describe
the temperature distribution, freezing front propagation and volume expansion during freezing. A 2D numerical
example was provided to demonstrate the ability of the proposed model in describing such process.
4 Acknowledgements
This work was funded by the Deutsche Forschungsgemeinschaft (DFG) through the project "Investigation and
calculation of frost heave considering specific boundary conditions of ground freezing". The funding is herewith
gratefully acknowledged.
References
[1] J. Bluhm, W.M. Bloßfeld, and T. Ricken. Energetic effects during phase transition under freezing-thawing
load in porous media - a continuum multiphase description and FE-simulation. ZAMM - J. Appl. Math.
Mech. / Zeitschrift für Angew. Math. und Mech. 94 (2014), pp. 586–608.
[2] W. J. Boettinger, J. A. Warren, C. Beckermann, and A. Karma. Phase-Field Simulation of Solidification.
Annu. Rev. Mater. Res. 32 (2002), pp. 163–194.
[3] B.D. Coleman and W. Noll. The thermodynamics of elastic materials with heat conduction and viscosity.
Arch. Ration. Mech. Anal. 13.1 (1963), pp. 167–178.
[4] W Ehlers. On thermodynamics of elasto-plastic porous media. Arch. Mech. 41.1 (1989), pp. 73–93.
94
[5] W. Ehlers and K. Häberle. Interfacial Mass Transfer During Gas–Liquid Phase Change in Deformable
Porous Media with Heat Transfer. Transp. Porous Media 114.2 (2016), pp. 525–556.
[6] A.G Fallis. Advances in Extended and Multifield Theories for Continua. Ed. by Bernd Markert. Vol. 59.
Berlin, Heidelberg: Springer, 2011, pp. 1689–1699.
[7] Y. Heider and B. Markert. A phase-field modeling approach of hydraulic fracture in saturated porous media.
Mech. Res. Commun. 80 (2017), pp. 38–46.
[8] D. M. Anderson, G. McFadden, and A. Wheeler. Diffuse-Interface Methods in Fluid Mechanics. Annu. Rev.
Fluid Mech. 30 (1997).
[9] B. Markert. A constitutive approach to 3-d nonlinear fluid flow through finite deformable porous continua.
Transp. Porous Media 70 (2007), pp. 427–450.
[10] I. Steinbach. Phase-field models in materials science. Model. Simul. Mater. Sci. Eng. 17 (2009), p. 073001.
[11] A. H. Sweidan, Y. Heider, and B. Markert. Modeling of PCM-based enhanced latent heat storage systems
using a phase-field-porous media approach. Contin. Mech. Thermodyn. (2019).
[12] H. Ziegler and C. Wehrli. The Derivation of Constitutive Relations from the Free Energy and the Dissipation
Function. Adv. Appl. Mech. 25 (1987), pp. 183–238.
95
96

Computational modelling of creep and rate dependent

strength of freezing soils
Rodolfo J. Williams M.
∗ 1 and Günther Meschke
1
1 Institute for Structural Mechanics, Ruhr University Bochum, Bochum, Germany
Abstract. In this work, the rate dependent eects of freezing soils such as creep and rate-dependent strength
are studied through a combination of viscoelastic and viscoplastic material models. The viscoelastic eective
properties are obtained through the upscaling of the elastic and viscous properties of the ice, soil, water material
phases. On the other hand, the viscoplastic component is phenomenological described with an overstress type
model. The proposed strategy is validated by comparing numerical predictions to experimental results.
Introduction
The rate-dependent behavior of freezing soils is strongly related to the microstructure of the ice crystal
phase, where the volumetric ice content and its cryostructure play an important role in the overall
macroscopic response. Several studies have been carried out during the past decades to gain insight
into the rate dependent eects such as creep, stress relaxation and rate-dependent strength behavior
on the basis of experimental evidence and then formulate constitutive models for frozen soils capable
of reproducing these features. Figure 1 (a) shows the three stages that can be distinguished during
creep under constant stress conditions: primary, secondary and tertiary creep, which can be found on
various laboratory studies. Many of these studies have provided empirical equations or phenomenological
constitutive relations to describe the macroscopic behavior of the frozen soil. In this work, the extended
Barcelona Basic Model (BBM) [3] with strength homogenization [5] is extended to include creep and rate
dependent eects through a combination of viscoelastic and viscoplastic material models. The viscoelastic
component to account for the primary creep is described by the micromechanical model presented in [2] ,
where the overall macroscopic response is based on the viscous behaviour of ice. On the other hand, the
viscoplastic component is phenomenologically described by Duvuat and Lions overstress model [1], which
allows to include viscous irreversible eects after yielding. The structure of this extended abstract is as
follows: First, the rate independent constitutive model is introduced in section 1. Then, the extension of
the latter model is described in section 2. Finally, for the validation of the proposed strategy, numerical
predictions are compared with experimental results in section 3.
1 Rate independent constitutive model for freezing soils
1.1 Freezing state: extended Barcelona Basic Model upon strength upscaling
The framework of the extended BBM with homogenized strength properties obtained through two-step
upscaling [5], is introduced in this section. The total strain ε̇ is dened through the additive decomposition
ε̇ = ε̇e + ε̇p . The stress-strain relation is assumed to be independent of the volumetric eective stress
p0 . For this reason, the swelling index is adopted as κ = vpmin /khom , with pmin = 1 kPa, v = 1/(1 − φ)
the specic volume and φ the porosity. For the determination of the homogenized bulk khom and shear
µhom moduli, the Mori-Tanaka scheme for a matrix inclusion morphology is employed for the case of fully
saturated frozen soil.
σ̇ = Cehom : ε̇e = Cehom : (ε̇ − ε̇p ) with Cehom = 3khom Ivol + 2µhom Idev (1)
∗ Corresponding author: rodolfo.williamsmoises@rub.mail
97
The temperature-dependent yield surface in the freezing state is written as

!nn " #
0 0 q 1 p0 + Shom
II
[T ]
f [p , q, po , T ] = II [T ](p0 + S II [T ])
+ ln 0 II [T ]
≤0 (2)
Mhom hom ln[rr] poT + Shom
The tensile strength ST , eective pre-consolidation pressure p0oT and temperature dependent compression
index λT are dened as:
! λλ−κ
−κ
!
cII [T ] p0o T
ST = hom
II [T ]
, and p0oT = with λT = λ (1 − r) exp[−βSf (Tf − T )] + r (3)
Mhom patm
2 Creep and rate dependent constitutive model for freezing soils
2.1 Viscoelastic behaviour of frozen soils
The viscous behavior of the composite is strongly dependent on the micromechanics of ice, which is
described as diusional thermal creep process, where the temperature dependent creep compliance Jdev
c
(t)
is expressed by:
! " !#
t Q i 1 1
c
Jdev (t) = 1/µc + Jcv ln 1 + v with Jcv (T ) = Jcv exp − − (4)
τc R T T
1
Figure 2 shows the inuence of the liquid saturation χL = (1 + (∆T /∆Tch ) 1−m )−m on the creep
secondary
creep
primary
creep
tertirary
strain
creep
time
(a) (b)
Figure 1. (a) The three creep stages: primary, secondary and tertiary creep. (b) Viscoelastic homogenization step.
1.0 7 1e 8
6 without L
0.8 with L
5
0.6 4
Jcv[Pa 1]
]
3
L[
0.4
2
0.2 1
0.020 020 15 10 5 0 5
15 10 5 0 5 T[°C]
T[°C]
(a) (b)
Figure 2. (a) Liquid saturation during freezing χL . (b) Inuence of the temperature and liquid saturation on the
creep compliance of the ice crystal Jvc .
compliance of the ice crystal Jvc . The total strain and the corresponding strain-stress relation for the
viscoelastic case are dened as:
ε̇ = ε̇e + ε̇ve and σ̇ = Cehom : ε̇e = Cehom : (ε̇ − ε̇ve ) (5)
98
For the determination of the viscoelastic behavior of the frozen soils, the framework of continuum mi-
cromechanics is used. It is based on the so called correspondence principle, where the viscoelastic ho-
mogenized properties in the time domain are obtained by inverse the Laplace-Carson transformation,
Figure 1 (b).
Z t ∂Jhom
∂ε ∂σ [t−τ ] dσ
hom
= J[t−τ ] : + : dτ (6)
∂t ∂t
| {z } | −∞ ∂t dτ
e
{z }
ε̇ ε̇ve
" !#
1 hom 1 hom 1 1 1 1 t − τ
Jhom
[t−τ ] = Jvol Ivol + Jdev Idev = Ivol + v
+ (1 − χL )Jhom ln 1 + v Idev (7)
3 2 3 khom 2 µhom τhom
∂Jhom
[t−τ ]
hom
1 ∂Jdev 1 v
Jhom
= [t − τ ]Idev = v Idev (8)
∂t 2 ∂t 2 (t − τ + τhom )
2.2 Rate dependent extended Barcelona Basic Model for freezing soils incorporating viscoelastic
creep
The total strain rate is dened through the additive decomposition using viscoelastic and viscoplastic
strain rate components and the corresponding strain-stress relation:
ε̇ = ε̇e + ε̇ve + ε̇vp and σ̇ = Cehom : ε̇e = Cehom : (ε̇ − ε̇ve − ε̇vp ) (9)
In the Duvaut-Lions model the viscoplastic strain rate and the eective pre-consolidation pressure with
τvp as the viscoplastic relaxation time, are formulated as:
(Cehom )−1 : 1 0
ε̇vp = : (σ − σ∞ ) and ṗ0 o = − (p − p0o∞ ) (10)
τvp τvp o
where σ∞ and p0o∞ denote the stress and the eective pre-consolidation pressure obtained from the
corresponding converged viscoelastic-plastic return mapping algorithm.
3 Numerical results
3.1 Uniaxial unconned compression creep test
In this section, the proposed model is validated by numerical re-analysis of a compression creep laboratory
test. The test was perfomed by Orth [4] on frozen Karlsruhe medium sand a temperature T = 10◦ C.
The prepared specimen had a water content of 18%. The creep test was carried out under constant
temperature and constant stress σ1 . The material parameters used for the validation example are: the
porosity φ0 = 0.18, bulk freezing temperature Tf = 273 K , freezing entropy Σf = 1.2 MPa , characteristic
cooling ∆Tch = 1 ◦ C, pore size distribution parameter m = 0.7, bulk modulus kS = 2650 GPa kL = 1000
GPa , kC = 917 GPa, shear modulus µS = 37.5 GPa, µL = 1 MPa, µC = 3.4 GPa. Further viscoelastic
and viscoplastic parameters are listed in Table 1. Figure 3 shows the creep-induced axial strain vs time
for dierent applied constant stresses. Good agreement between numerical and experimental results for
stresses between σ1 = 3MPa to σ1 = 6MPa are found. However, for stress levels higher than σ1 = 6MPa
discrepancies in the results are observed.
Conclusions
The presented constitutive model has the capability of reproducing primary creep behavior based on the
information of the microstructure of the frozen soil and the secondary with the help of the overstress
model. Results show, however, the limitations of the model to predict tertiary creep, which might be
associated with damage processes in the ice matrix. This is the topic of a current extension of the model.
99
Table 1. Viscoelastic and viscoplastic material properties employed in the in the numerical simulation
Material parameter Symbol Numerical values Units
Creep compliance for crystal ice Jcv 13000 µm/m/MPa
Creep relaxation time for crystal ice τcv 1 s
Shape parameter nn 2
Spacing ratio rr 2
Swelling index κ 0.0002
Compression index λ 0.005
Parameter dening max. compressive stiness r 0.8
Parameter controlling stiness grow rate β 0.2
Preconsolidation p0o 1 MPa
Viscoplastic relaxation time 3 ◦ C to 6 ◦ C τvp 1000 s
Viscoplastic relaxation time 7 ◦ C to 10 ◦ C τvp 250 s
16
14
12
Pa
Pa
8M
1 = 10MPa
9M
1=
10
1=
experimental
1[%]
8 numerical
a
6 1=
7MP 1 = 6MPa
4 1 = 5MPa
2 1 = 3MPa
00 20000 40000 60000 80000 100000 120000

t[s]
Figure 3. Evolution of axial strain of frozen sand for dierent prescribed constant stress levels at T = 10 ◦ C.
Acknowledgments
Financial support was provided by the German Research Foundation (DFG) within the scope of Subpro-
ject C1 of the Collaborative Research Centre SFB 837.
References
[1] G. Duvaut and J.L. Lions. Les inéquations en mécanique et en physique. Travaux Recherches Math.
Paris: Dunod, 1972. url: http://cds.cern.ch/record/212412.
[2] R. Lackner, C. Pichler, and A. Kloiber. Articial Ground Freezing of Fully Saturated Soil: Vis-
coelastic Behavior. Journal of Engineering Mechanics 134.1 (2008), pp. 111. issn: 0733-9399. doi:
10.1061/(asce)0733-9399(2008)134:1(1).
[3] S. Nishimura, A. Gens, S. Olivella, and R. J. Jardine. THM-coupled nite element analysis of frozen
soil: formulation and application. Géotechnique 59.3 (2009), pp. 159171. issn: 0016-8505. doi: 10.
1680/geot.2009.59.3.159. url: http://www.icevirtuallibrary.com/doi/10.1680/geot.
2009.59.3.159.
[4] W. Orth. Gefrorener sand als werksto: Elementversuche und materialmodell. Vol. 100. Institut für
Bodenmechanik und Felsmechanik der Universität Fridericiana in . . ., 1986.
[5] MM. Zhou and G. Meschke. A multiscale homogenization model for strength predictions of fully and
partially frozen soils. Acta Geotechnica 13.1 (2018), pp. 175193. issn: 18611133. doi: 10.1007/
s11440-017-0538-0.
100
MS 06: Modeling and simulation of fatigue in brittle and ductile materials

The FKM guidelines: Recommendations for fatigue life

calculations of industrial components
Melanie Fiedler ∗ 1 and Michael Vormwald2

1 TU Dresden, Germany
2 TU Darmstadt, Germany
Abstract. The FKM guideline Rechnerischer Festigkeitsnachweis [8] is widely used in the german industry
to calculate the fracture fatigue lives of components made of steel or aluminium alloy under constant or vari-
able amplitude loadings. The guideline is based on a simple approach for Woehlercurves: constant slope and
constant kink point for all kinds of ductile steels. The fatigue strength of the component can be approxi-
mated by the ultimate tensile strength and the geometrical properties of the component. No other material
parameters are required. The fatigue lives for components under variable amplitude loadings can be calcu-
lated by elementary or consequent Miner rule. The elastic-plastic material behavior is neglected when using
the FKM guideline also no crack growth simulation is required. In 2019 the new FKM guideline nonlinear
[4] was published based on the Local Strain Approach. This new guideline considers elastic-plastic material
behavior in a fatigue life calculation for crack initiation. The cyclic material behavior and the Woehlercurves
for crack initiation can also be approximated by the ultimate tensile strength and the damage accumulation
is also based on elementary or consequent Miner rules. In [3] the U-Concept is published, which is a new
add-on to the Local Strain Approach. The U-Concept is applicable for components of steel under constant
or variable amplitude loadings and transfers the advantages of the Local Strain Approach from crack ini-
tiation to fracture. This presentation gives an overview about the new add-on and presents experimental
versus calculated results for a large database including more than 1500 experiments for notched compo-
nents under constant amplitude loading and more than 1400 experiments under variable amplitude loading.
Introduction
The Research Associations for Mechanical Engineering (FKM) is publisher of three dierent guidelines for
structural durability. The new FKM guideline nonlinear (FKM nonlinear, section 2) [4] for calculations
with failure criteria technical crack initiation 0.25 mm ≤a≤3 mm, the fracture mechanical guideline
for crack growth simulations from crack initiation to fracture and the Rechnerischer Festigkeitsnachweis
(FKM linear, section 1) [8] for failure criteria fracture. The FKM linear considers elastic material behavior
in the fatigue life calculation while the FKM nonlinear is able to consider elastic-plastic material behavior
by using the Local Strain Approach (LSA). In section 3 the U-Concept is described, which combines the
LSA with the failure criteria fracture. The results of the U-Concept in comparison to FKM linear are
discussed in section 4 for constant and variable amplitude loadings for a probability of failure P = 50%.
1 The FKM guideline linear
This paper considers the local stress approach proposed by the FKM guideline linear [8]. For a local
stress amplitude σa = Kt · Sa the approach predicts the fracture fatigue life NCalc,frac of a structure under
constant amplitude loading:

 σW ·KAK,σ /KW K,σ k
σa for σa > σW · KAK,σ /KW K,σ
NCalc,frac = ND · (1)
∞ for σa ≤ σW · KAK,σ /KW K,σ
with the fatigue limit ND = 106 and the slope k =5 for steel, the endurance strength of the material
σW , the mean stress eect factor KAK,σ and the component factor KW K,σ .
∗ Corresponding author: melanie_edler@tu-dresden.de ORCID iD: ORCID (optional)
101
The endurance strength σW for tension/compression can be approximated by the ultimate tensile strength
Rm using σW = fW,σ · Rm with the tension-compression-factor fW,σ = [0.4; 0.45] for steel structures.
For structures under variable amplitude loading the FKM guideline linear uses the elementary or conse-
quent miner rule for damage accumulation. For damage sum D=1 the approach of the linear guideline
tends to overestimate the fracture fatigue live of a component under variable amplitude loading in com-
parison to experimental results, see [5]. Based on an approach by [5], eective damage sums 0.3 ≤ De ≤ 1
have to be taken into account during the calculation to guaranty conservative calculated fatigue lives.
2 The FKM guideline nonlinear
The FKM guideline nonlinear [4] is based on the Local Strain Approach (LSA) and is able to consider
elastic-plastic material behavior during the fatigue life calculation for crack initiation. Based on an elastic
FE-calculation, elastic-plastic material behavior can be assumed by the ow curve approach by Neuber
[7] or by Seeger and Beste [10]. The elastic-plastic material behavior is described by the cyclic stabilized
Ramberg-Osgood-equation:
σa σa 1/n
0
εa = + (2)
E K0
The material parameters E, K 0 and n0 can be approximated by the FKM method [13] based on the
ultimate tensile strength Rm and the material group. The local stress strain path is simulated with a
Hysteresis Counting Method (HCM) [1], which combines the rainow counting method with the elastic
plastic material behavior and is able to consider Masing's law and the three memory eects. The guideline
considers two damage parameters. Damage parameter PRAM which is based on PSWT [11] and PRAJ based
on PJ by [12]. Both damage parameters are able to consider mean stress eects, PRAJ is also able to
take sequence eects into account. The damage parameter Woehlercurves for crack initiation can be
approximated by
(
(PRAM /PRAM,Z )1/d1 for PRAM > PRAM,Z
NCalc,CI = 103 · (3)
(PRAM /PRAM,Z )1/d2 for PRAM ≤ PRAM,Z
and
(
(PRAJ /PRAJ,Z )1/d for PRAJ > PRAJ,D
NCalc,CI = (4)
∞ for PRAJ ≤ PRAJ,D
with grid points PRAM,Z and PRAJ,Z , fatigue strength PRAJ,D and slopes d, d1 , d2 which are approximated
by the ultimate tensile strength Rm , the material group, [13], and the component factors fRAM and fRAJ ,
[4]. The damage accumulation for variable amplitude loadings is also based on elementary or consequent
Miner rules, see [4].
3 The U-Concept
The U-Concept (UC) by [3] is a simple add-on to the LSA in order to calculate technical fracture
fatigue lives NCalc,frac for non-welded structures of steel based on elastic-plastic material behavior. The
UC contains dierent approaches for damage parameters PRAM and PRAJ . The approaches PRAM,KP
and PRAJ,KP propose a constant shift of the damage parameter Woehlercurves from crack initiation to
fracture. The shift factor f depends on the ultimate tensile strength Rm , the plastic load factor Kp and
the nominal stress ratio R, see [3].
NCalc,frac = f (Rm , Kp , R) · NCalc,CI (5)
The approach PRAJ,U considers the dierent slopes of crack initiation and fracture Woehlercurves which
can be observed in experiments, [9]. The shift function fU depends on the elastic-plastic calculated stress
amplitude σa and mean stress σm of the closed hysteresis loop as well as the ultimate tensile strength
Rm , the plastic load factor Kp and the nominal stress ratio R, [3].
NCalc,frac = fU (σa , σm , Rm , Kp , R) · NCalc,CI (6)
The UC considers the eective damage sum approach by [5] in combination with a method correction
factor fmeth to De,NL = fmeth · De with fmeth = 1 for FKM linear. The damage accumulation for
variable amplitude loadings is also based on elementary or consequent Miner rules.
102
4 Results
The experiments of this section are mostly taken out of the database DaBef, [2]. The structural database
consists of experimental results for notched specimens under constant or variable amplitude loadings for
steel, cast steel and aluminum alloy. For constant amplitude loadings, the database was extended by the
sources [6, 9].
4.1 Constant amplitude loadings
In gure 1 results for notched specimens under constant amplitude loadings are shown. On the left side,
the results of the U-Concept with approach PRAJ,U are compared to the experimental results, on the
right side, the results for FKM linear are plotted against the experimental fatigue lives. The U-Concept
shows a smaller scatter of the quotient qN = NExp,frac /NCalc,frac than the approach by FKM linear.
Figure 1. Comparison of experiments for notched specimens under constant amplitude loading with calculated
fracture fatigue lives with approaches FKM linear (right) and U-Concept PRAJ,U (left) with P = 50%,
[3]
4.2 Variable amplitude loadings
Figure 2 (left) shows a comparison between the experimental results by [6] for specimens under vari-
able amplitude loading (bending, R̄ = −1) with calculated results. The loading sequence is normal
distributed and the calculated results are neither calculated with the approaches of the U-Concept
(PRAM,KP , PRAJ,KP and PRAJ,U ) or with the approach of the FKM linear for a probability of P = 50%.
In combination with the consequent miner rule the approaches PRAJ,KP and PRAJ,U show the best match
with the progression of the experimental results. Figure 2 (right) shows the results of the quotient of
experimental to calculated fatigue lives qN = NExp,frac /NCalc,frac for all concepts. The 1413 experiments
and the variable amplitude loading sequences were taken out of the database DaBef, [2].
103
Figure 2. (left) Comparison of experiments by [6] with calculated, fracture fatigue live curves for the dierent
concepts with P = 50% and D = De,NL , [3]; (right) Comparison of the accuracy of the dierent
approaches with 1413 experiments [2] of specimens under variable amplitude loading for life quotient
qN = NExp,frac /NCalc,frac and D = De,NL , [3]. The accuracy is described by mean value mN and life
variance TN , assuming a logarithmic normal distribution.
References
[1] U. Clormann. Örtliche Beanspruchungen von Schweiÿverbindungen als Grundlage des Schwingfes-
tigkeitsnachweises. Dissertation. Darmstadt: TU Darmstadt, 1986.
[2] F. Ellmer, K. Hinkelmann, K.-G. Eulitz, and Esderts A. Datenbank und Auswertesystem Betriebs-
festigkeit. FKM-report No. 313. 2011.
[3] M. Fiedler. Das U-Konzept - Technische Bruchlebensdauern auf Basis des Örtlichen Konzepts.
Dissertation. Darmstadt: TU Darmstadt, submitted 2019, unpublished.
[4] M. Fiedler, M. Wächter, I. Varfolomeev, M. Vormwald, and A. Esderts. Richtlinie nichtlinear - Rech-
nerischer Festigkeitsnachweis unter expliziter Erfassung nichtlinearen Werkstoverformungsverhal-
tens. VDMA Verlag, Frankfurt am Main, 2019.
[5] K. Hinkelmann. Korrekturfunktionen zur Verbesserung der rechnerischen Lebensdauerabschätzung
bei schwingender Beanspruchung. Dissertation. TU Clausthal, Clausthal-Zellerfeld, 2012.
[6] M. Hück and J. Bergmann. Bewertung der Schwingfestigkeit der mikrolegierten Stähle 27 Mn VS 6
und 38 Mn VS 5 - Teil II. FKM-report No. 122. 1992.
[7] H. Neuber. Theory of Stress Concentration for Shear-Strained Prismatical Bodies with Arbitrary
Nonlinear Stress-Strain Law. Trans. ASME, J. Appl. Mech. 28 (1961), 544550.
[8] R. Rennert, E. Kullig, M. Vormwald, A. Esderts, and D. Siegele. Rechnerischer Festigkeitsnachweis
für Maschinenbauteile aus Stahl, Eisenguss- und Aluminiumwerkstoen. VDMA-Verlag, Frankfurt
am Main, 6. revised ed., 2012.
[9] H. Saal. Einuÿ von Formzahl und Spannungsverhältnis auf die Zeit- und Dauerfestigkeit und Riÿ-
fortschreitungen bei Flachstäben aus St52, Report 17. Veröentlichung des Instituts für Statik und
Stahlbau der Technischen Hochschule Darmstadt, 1971.
[10] T. Seeger and A. Beste. Zur Weiterentwicklung von Näherungsformeln für die Berechnung von
Kerbbeanspruchungen im elastisch-plastischen Bereich. Bruchvorgänge und Schadensanalyse (1977).
Ed. by Bruchvorgänge und Schadensanalyse, 156.
[11] R.N. Smith, P. Watson, and T.H. Topper. A stress-strain function for the fatigue of metals. Journal
of Materials 5 (1970), pp. 767778.
[12] M. Vormwald. Anriÿlebensdauervorhersage auf der Basis der Schwingbruchmechanik kurzer Risse.
Veröentlichung des Instituts für Stahlbau und Werkstomechanik der TH-Darmstadt, Vol. 47,
1989.
[13] M. Wächter. Zur Ermittlung von zyklischen Werkstokennwerten und Schädigungsparameterwöh-

lerlinien. Dissertation. Clausthal-Zellerfeld: TU Clausthal, 2016.
104

An Approach to Characterize Fatigue Crack growth by

Material Forces
Jad Khodor1 and Michael Kaliske ∗ 1
1 Institute of Structural Analysis, Technische Universität Dresden, 01062 Dresden, Germany
Abstract. A new approach is proposed to model fatigue crack growth using material forces based on the
degradation of the fracture toughness. The variable used in the degradation function is the accumulated norm
of the material force acting on the crack tip. A non-local accumulation is performed in order to avoid mesh
dependency.
Introduction
Nowadays, many structures are subjected to cyclic loading, where the maximum value of the load could be
much smaller than the ultimate design load. Although the applied load is smaller than the ultimate one,
some structures fail after a certain number of cycles. In the framework of fracture mechanics the crack
propagation can be divided into three main stages. In the first stage there will be nucleation of micro-
cracks that merge together to form a macro-crack. The second stage, describes the stable propagation of
the macro-crack, whereas in the third stage the crack propagates in an unstable manner. The pioneering
law in describing the second stage is known as Paris’ law [5] that correlates the stress intensity factor to
the rate of crack length per cycle. However, the use of the stress intensity factor is only limited to cases
where the theory of linear elastic fracture mechanics (LEFM) is governing. In the work of [2], the cyclic
J-integral ∆J is introduced for gross plasticity to overcome the limitation of the stress intensity factor. In
this paper, a new approach to characterize fatigue crack growth using material forces or configurational
forces will be presented.
1 Material Forces
Configurational forces or material forces are non-Newtonian forces which act on an inhomogeneity
within a body. This inhomogeneity can be an inclusion or a crack tip. The material force that acts on
the crack tip is considered as the crack driving force. The direction of the material force vector at the crack
tip can be used to determine the crack propagation direction, whereas its norm can be used to calculate
the energy release rate. The material force can be computed using Eshelby’s thought experiment [3].
Eshelby considered an inclusion within a homogeneous elastic body. The body is subjected to traction
forces as shown in Figure 1. He studied the energy variation due to an infinitesimal displacement δX of
the surface P to the new position P 0 . The energy variation is defined as the dot product between the
material force acting on the inclusion and the infinitesimal displacement and can be written as
δψ = F · δX, (1)
where F is the material force that is defined as

Z
F = − [ψI − ∇X ut σ]ndS, (2)
P
where ψ is the strain energy, I is the second order identity tensor, σ is the Cauchy stress tensor and
∇X u is the gradient of the displacement field. For a small strain linear elastic material, the Eshelby
∗ Corresponding author: Michael.kaliske@tu-dresden.de
105
Figure 1. Eshelby’s thought experiment
stress tensor is defined as Σ = ψI − ∇X ut σ. In case of finite strains, the Eshelby stress tensor is
defined as Σ = ψI − F t P , where F and P are the deformation gradient and the Piola-Kichhoff stress
tensor, respectively.
2 Fatigue Crack Growth Using Material Forces
2.1 Fracture Toughness Degradation
In the work of [1], a degradation function has been introduced to reduce the fracture toughness using the
phase-field method. The accumulated variable used in the degradation function was the active part of the
strain energy density. In this study, the norm of the material force at the crack tip will be accumulated.
The degradation function is defined as
0 ≤ f (γ(t)) ≤ 1, (3)
(
1 if γ(t) ≤ γ0
f (γ(t)) = 2γ0 , (4)
( γ(t)+γ 0
)2 if γ(t) ≥ γ0
where γ(t) is the accumulated variable and γ0 is a material parameter. The degradation variable is
accumulated as
Z t
γ(t) = H(γ γ̇)|γ̇|dτ, (5)
0

where H(γ γ̇) is the Heaviside function and γ = F crack is the norm of the material force acting on the
crack tip. The numerical accumulation of the material force norm is obtained as
γ n+1 = γ n + ∆γ, (6)
γn+1 − γn
∆γ = |γn+1 − γn | H( ). (7)
∆t
Griffith [4] states that the condition for crack propagation is defined as
G − Gc = 0, (8)
where G and Gc are the potential energy release rate and the fracture toughness, respectively. In case of
fatigue crack propagation, the fracture toughness is degraded and therefore Eq. (8) can be written as
G − f (γ(t))Gc = 0. (9)
2.2 Non-Local Accumulation of The Degradation Variable
In order to avoid mesh dependency of the degradation function, the accumulation should be achieved in
a non-local way. ∆αcracktip and ∆αnode represent the accumulation increment at the crack tip node and
at a node next to the crack tip, respectively. The non-local accumulation over the nodes surrounding the
crack tip is defined as
2
−
kxcracktip −xnode k
∆αnode = ∆αcracktip × e 2l2
c , (10)
where lc is a length scale parameter, xcracktip and xnode are the position vectors of crack tip node and
the surrounding node, respectively.
106
3 Numerical Example
The validation of this approach is achieved by applying a cyclic load on a tensile test specimen with an
initial crack length a0 = 0.2mm. The dimensions and the boundary conditions of the specimen are shown
in Figure 2. The specimen is subjected to prescribed displacements at the boundaries with a maximum
value umax = 0.2 mm and loading ratio R = uumaxmin
= 0. A linear elastic isotropic material, which uses a
Young’s modulus E = 10GPa and a Poisson’s ratio ν = 0.25, has been utilized in this simulation.
8 mm
2 mm
a0
Figure 2. Tensile test specimen
The fracture parameters used are Gc = 150N mm−1 , γ0 = 3000N mm−1 and lc = 0.2mm. The simulation
is performed with 2-D 4 node quadrilateral elements at plane strain conditions. The results obtained
from the simulation are the number of cycles and material force norms corresponding to each crack
length. Using these data the crack increment per cycle can be plotted against the material force norm as
represented in Figure 3.
Figure 3. Double logarithmic plot of the crack increment per cycle and the material force norm at the crack tip
(small strains)
It can be seen from Figure 3 that the three stages of fatigue crack growth are depicted by the proposed
approach, the Paris’ parameters can be obtained from the second stage that is located between the
two dotted lines. The relation between the crack increment per cycle and the material force norm is
1.89
−6 crack
da
dn = 10 F . The idea behind using material forces is to overcome the limitations of the stress
intensity factor. In case of straight crack propagation, the material force norm is equal to the J-integral.
In this example, the crack propagates straightly without kinking. Therefore, the relation between the
stress intensity factor and material force norm is defined as

crack K 2
F = 0, (11)
E
where K is the stress intensity factor, E 0 = E and E 0 = 1−ν
E
2 for plane stress and plane strain conditions,
respectively. A Neo-Hookean material model is used to validate the approach in case of finite strain
theory, which has the following energy function
1 2 1
ψ= (K0 − G)(J − 1)2 + G(Ic − 3 − 2ln(J)), (12)
2 3 2
107
where J is the determinant of the deformation gradient and Ic = tr(F t F ). The parameters K0 and G
are computed using E and ν, where
E
K0 = , (13)
3(1 − 2ν)
E
G= . (14)
2(1 + ν)
The same values for E and ν are used for the hyperelastic case. The plot of the crack increment per cycle
and the material force norm is shown in Figure 4.
Figure 4. Double logarithmic plot of the crack increment per cycle and the material force norm at the crack tip
(finite strains)
It can be seen from Figure 4 that the three stages of fatigue crack growth are depicted for the finite
strain theory case.
The relation between the crack increment per cycle and the material force norm is
2.13
crack
da
dn = 5 × 10
−7
F .
4 Conclusion
The proposed approach is able to depict the three stages of fatigue crack propagation. The newly
introduced material parameters lc and γ0 can be calibrated with some experiments in order to get the
same Paris’s parameters, as of the experiments. The use of material forces in fatigue crack growth helps
in overcoming the limitations of the stress intensity factor.
Acknowledgments
The authors of this work gratefully acknowledge ANSYS Inc., Canonsburg for their financial support.
References
[1] P. Carrara, M. Ambati, R. Alessi, and L. De Lorenzis. A Novel Framework to Model The Fatigue
Behavior of Brittle Materials Based on a Variational Phase-Field Approach (2018).
[2] N. Dowling and J. Begley. Fatigue Crack Growth During Gross Plasticity and the J-Integral. ASTM
STP 590 (1976), pp. 82–103.
[3] J.D. Eshelby. The Force on an Elastic Singularity. Philosophical Transactions of the Royal Society
of London 244 (1951), pp. 87–112.
[4] A. Griffith. The Phenomena of Rupture and Flow in Solids. Philosophical Transactions of the Royal
Society of London, Series A 221 (1920), pp. 163–198.
[5] P. Paris and F. Erdogan. A Critical Analysis of Crack Propagation Laws. Journal of Basic Engi-
neering 85 (1963), pp. 528–533.
108

Extension of the notch strain concept to transient cyclic

material behaviour and residual stresses
1
David Kühne ∗
, Martha Seiler2 , Franz Ellmer1 , Peter Hantschke1 and Markus Kästner2
1 Structural Durability Group, TU Dresden, Dresden, Germany
2 Chair of Computational and Experimental Solid Mechanics, TU Dresden, Dresden, Germany
Abstract. A number of different forming processes result in significant residual stresses within the produced
components, which lead to changes of their fatigue life properties. Furthermore materials are able to change
their cyclic hardening behaviour during fatigue life which can also result in residual stresses. While state-of-
the-art fatigue concepts (e.g. the notch strain concept) assume stabilized cyclic material behaviour throughout
the entire fatigue life of the component, this assumption is not adequate for residual stress- and cyclic hardening
afflicted components. Hence, a generalization to transient cyclic material behaviour with an asymmetry in
tension and compression is introduced. To consider the load sequence order for variable amplitude load series
on the fatigue life, the hysteresis counting method (HCM) is used. The damaging effect of each half-hysteresis
is evaluated by a damage parameter according to Smith-Watson-Topper based on a strain Wöhler curve which
represents the second set of input data. The fatigue concept is applied to the analysis of residual stress afflicted
components.
Introduction
Forming processes generally lead to residual stresses. Compressive residual stresses can have a beneficial
impact on the fatigue life of components while tensile residual stresses cause a shortening in fatigue life.
Reiß and Merklein et al. [2, 11] demonstrated that different forming processes induce residual stresses
into a component, which have a significant impact on its fatigue life. The Mechanical fatigue describes
the process of crack initiation from the undamaged material on during the occurence of cyclic loads.
From this point, stabilized crack growth occurs until a critical crack length is reached and the residual
crack initiates. The fatigue life is defined here as the number of cycles until the crack initiates. This
process is characterized by microscopic dislocation movements and discontinuities at the crystall grain
scale. These phenomena can be summarized to an elastic-plastic-material behaviour on the macroscopic
scale. The most common material concept considering fatigue is the notch strain concept. In general
it assumes unchanging cyclic material behaviour throughout the entire fatigue life of the component.
This assumption is not adequate for transient and residual stress afflicted material behaviour. As will be
demonstrated later, residual stresses in combination with transient material behaviour behaviour have a
considerable influence on the fatigue life. Therefore we extend the standard notch strain concept to a
transient approach. The extended fatigue concept is depicted in Fig. 1. The circled numbers denotes
the different features of the concept, described in the section below.
1.1 Standard Notch Strain Concept
The notch strain concept [9] is a material-based concept which retraces a representative local stress-strain
path and calculates the damage associated with it. The local stress strain path generally depends on
the notch geometry, the elasto-plastic material parameters and the external load-time sequence. It is
reduced to a series of reversal points. Thus, the external load data for the fatigue-analysis is a sequence
∗ Corresponding author: david.kuehne@tu-dresden.com
109
of elastically calculated stresses at the considered material point over the loading path. The state of the
art here is to convert the linear elastically calculated strain states using the Neuber rule
6 [7].
As an input for the revaluation serves the sequence of elastic stresses and strains, which is derived by
scaling the stress-strain state of one elastic baseline simulation 2 to all reversal points of the loading
path 4 . Result of the revaluation is an elasto-plastic stress-strain path from which hystereses can be
identified. The shape of the initial stress-strain path is determined by the cyclic stress strain curve
(CSSC) described by the Ramberg-Osgood Equation [10], where the elastic stress-strain state σel , εel is
revaluated to elasto-plastic stress and strain amplitudes σa , εa
1/n0
σa σa 1/n
0
∆σ ∆σ
εa = εa,el + εa,pl = + , (1) ∆ε = +2 . (2)
E K0 E 2K 0
Here, K 0 and n0 are material constants to describe hardening behaviour. The CSSC can be obtained from
either constant amplitude tests (CA-Tests) or incremental step tests (IST). Although the stress-strain
path changes with increasing damage, it is common practise to use the stabilized CSSC which manifests
after 50 percent of fatigue life. On top of that, the CSSC is usually symmetrised, neglecting the different
behaviour in tension and compression.
The Masing behaviour [5] considers the hysteresis shape which corresponds to the CSSC scaled by the
factor two in the direction of stress- and strain. Hereby, the stress and strain path is described by relative
values ∆σ, ∆ε with regard to the reversal points. The loading path is processed step by step and can
thus take into account effects of the chronological loading order. To consider variable load amplitudes or
random loads, the hysteresis counting method (HCM) [1] is used to determine the memory behaviour.
The fatigue strength of the material is represented by the strain-life curve 5 . It relates the total strain
amplitude εa with the endurable number of load cycles N and is described using the Manson-Coffin-
Morrow approach [4]
σf0
εa = εa,el + εa,pl = · (2N )b + ε0f · (2N )c (3)
E
with the parameters σf0 , b determining the contribution of the elastic strain and ε0f , c the one of the
plastic strain part. The damaging peffect of each hysteresis including its mean stress value is evaluated
by a damage parameter PSWT = (σa + σm ) · εa · E (P-value), e.g. according to Smith-Watson-Topper
[12]. With the P-value 8 the strain-life curve is converted into a damage-life curve, which provides the
number of load cycles N , that the material could theoretically endure, if it was loaded with hystereses
with that specific P-value only. According to the linear damage accumulation hypothesis [6, 8] the total
damage
Xn
1
D= (4)
i=1
Ni
can be summed up from all n hystereses with the damage contributions 1/Ni . The total damage charac-
terises the resistance of the component against the applied load: if D reaches 1, failure occurs theoretically,
i.e. a crack initiation can be observed. With this criterion, the life time can be predicted from the given
load sequence 4 .
1.2 Extension of the Fatigue Concept
As described in the publication [3] the Notch Strain Concept is extended to asymmetric and transient
cyclic material behaviour. For the asymmetric behaviour the material law (CSSC) in the standard notch
strain concept must distinguish between tensile and compressive stresses. Therefore, an asymmetric
stress-strain relationship is introduced by applying different cyclic hardening parameters K+
0
and K−0
and n+ and n− for tensile and compressive stresses, which must be related to the damage variable D to
0 0
describe the transient behaviour over the course of fatigue life

3
( (
0
0 K+ (D) σ ≥ 0 0 n0+ (D) σ ≥ 0
K → 0
n → . (5)
K− (D) σ < 0 n0− (D) σ < 0
110
Hence, the Ramberg-Osgood (1) equation including Masing behaviour (2) can be rewritten as
1/n0± (D) !
|∆σ| |∆σ|
∆ε± (∆σ, γ) = sign(∆σ) + γ± 0 (D) . (6)
E K±
For an asymmetric stress-strain relationship, the hysteresis curve must be adjusted with the parameter
γ to fit the experimental CSSC. Therefore, the parameters γ+ and γ− are introduced to describe the
asymmetric reloading curves for tension and compression.
Due to the symmetrized CSSC Masing behaviour can be assumed. Thus the absolute value of hysteresis
paths despite loading and unloading paths are equal and hysteresis are compulsory closed at their initial
point. While in the non-transient approach no hysteresis shift occurs, the transient approach 7 causes
a shift in stress- or strain-direction. To consider residual stresses , the initial point of the stress strain
1
path is shifted along the initial CSSC. Thus the the first point of the stress-strain-path starts at σm with
the related εm wich is associated to the CSSC. Throughout the shift, the hysteresis closure no longer
necessarily takes place at the starting point of the hysteresis and the determination of the damage is
not trivial anymore. Hence the evaluation with the damage parameter is no longer performed for closed
hystereses but for half hystereses 9 . Hence, the linear damage accumulation hypothesis, with 0.5/Ni
being the damage contribution of one half hysteresis, is modified to
n
X 0.5
D= . (7)
i=1
Ni
Figure 1. Scheme of the extended notch strain concept.
2 Demonstration Example
The concept has been validated and applied to analyse the general influence of residual stresses on the
transient cyclic behaviour. To do so, moderate residual stresses σres ∈ {−300, 300} MPa are included
111
in the analysis as mean stress σm . The superimposed residual stresses shift, the hysteresis curves to
different stress levels. Higher stresses lead to a faster damage accumulation and, as a consequence of
the asymmetric transient cyclic behaviour, the hystereses become tilted with increasing fatigue damage.
In the next step, the influence of the transient behaviour of the CSSC on fatigue life 10 is analysed in
terms of the cyclic hardening parameter K 0 for different residual stresses. Fig. 2 depicts three different
cyclic characteristics, K+1
0
as it was fitted to the experimental data, obtained from an Incremental Step
Test with an unnotched specimen and additionally two functions K+2 0
and K+30
with less pronounced
cyclic softening, for comparison. Furthermore, the estimated fatigue life Nf for the three transient K 0 in
dependence of the residual stresses is also shown. Moreover, a constant K 0 = K+10
(D = 0.5), which allows
for a comparison with the standard concept, is considered. Remarkable is the deviation of the fatigue
lifes at very high compressive residual stresses. It follows that the progression of the cyclic material
parameters in combination is especially relevant for fatigue life in case of compressive residual stresses.
Figure 2. Comparison of the fatige life for different residual stresses, calculated for a strain amplitude of εa = 0.5%
with the standard notch strain concept and the extended one.
References
[1] U.H. Clormann and T. Seeger. Rainflow-HCM: ein Hysteresisschleifen-Zählalgorithmus auf werk-
stoffmechanischer Grundlage. Fachgebiet Werkstoffmechanik, Technische Hochschule Darmstadt.
[2] Influence of machining process on residual stresses in the surface of cemented carbides. Procedia
Engineering 19 (2011). 1st CIRP Conference on Surface Integrity (CSI), pp. 252 –257.
[3] D. Kühne, C. Guilleaume, M. Seiler, P. Hantschke, F. Ellmer, T. Linse, A. Brosius, and M. Kästner.
Fatigue analysis of rolled components considering transient cyclic material behaviour and residual
stresses. Production Engineering 13.2 (2019), pp. 189–200.
[4] SS Manson. Fatigue: a complex subject—some simple approximations. Experimental mechanics 5.7
(1965), pp. 193–226.
[5] G. Masing and W. Mauksch. Eigenspannungen und Verfestigung des plastisch gedehnten und ges-
tauchten Messings. Berlin, Heidelberg: Springer Berlin Heidelberg, 1925, pp. 244–256.
[6] M. A. Miner. Cumulative damage in fatigue. Journal of applied mechanics. Band 12, Nr. 3, S.
159–164. Journal of applied mechanics., 1945.
[7] H. Neuber. Theory of stress concentration for shear-strained prismatical bodies with arbitrary
nonlinear stress-strain law. Journal of applied mechanics 28.4 (1961), pp. 544–550.
[8] A. Palmgren. Die Lebensdauer von Kugellagern. Zeitschrift des Vereins Deutscher Ingenieure. Band
68, Nr. 14, 1924, S. 339–341. Zeitschrift des Vereins Deutscher Ingenieure., 1924.
[9] D. Radaj, C. M. Sonsino, and W. Fricke. Fatigue assessment of welded joints by local approaches.
Woodhead publishing, 2006.
[10] W. Ramberg and W. R. Osgood. Description of stress-strain curves by three parameters (1943).
[11] A. Reiss, U. Engel, and M. Merklein. Investigation on the Influence of Manufacturing Parameters
on the Fatigue Strength of Components. In: The Current State-of-the-Art on Material Forming.
Vol. 554. Key Engineering Materials. Trans Tech Publications Ltd, July 2013, pp. 280–286.
[12] K.N. Smith, T.H. Topper, and P. Watson. A Stress-strain Function for the Fatigue of Metals.
112

Transferability concepts of Fatigue Strength for high

purity copper components in electrical engines
1
Stefan Pruy , Markus Kästner
2 and Peter Hantschke
2
1
ZF Friedrichshafen AG, Germany
2
TU Dresden, Germany
Abstract.
The fatigue strength of high purity copper components in electrical engines is inuenced by the
manufacturing process and the geometrical properties of the notches. A test setup on an electro-
dynamic shaker allows to get component specic S-N curves. The manufacturing inuence can
either be evaluated by hardness measurements or a process simulation, which is described for a
forming process in this study. A comparison of common methods for the geometrical transfer-
ability shows the untypical behaviour of high purity copper. Neither the relative stress gradient
nor the high stress volume indicate a correlation to the fatigue strength for the given specimens.
Introduction
The electrication of the powertrain in vehicle construction leads to new requirements for the
operationally stable and reliable design of components and assemblies. For a valid design process
according to vibrational and time dependent temperature loads a customized fatigue life evalua-
tion is required. This is especially important for the materials used in mechatronic assemblies as
they typically react sensitive to environmental and operational loads. Within the scope of this
study, the inuences on the fatigue strength of high purity copper components are identied. A
concept for the transferability of the fatigue strength which considers the manufacturing process
and the geometrical inuence on the notch stresses is proposed.
1 Inuences on the fatigue strength

Inuences on the fatigue strength for electrical connectors can be divided into two parts: the
manufacturing processes and the geometrical inuences which are coupled with the load situation.
1.1 Manufacturing processes

To produce economical and reliable busbars with high eciency, high-purity copper sheets or
wires commonly serve as raw materials. The following manufacturing processes are typical and
inuence the fatigue strength:
• Rolling: Rolling of the raw copper material allows to produce a specic sheet thickness
and strength properties.
• Stamping and cutting: Stamping and cutting allows to get specic semi-nished parts
out of the metal sheets or wires.
113
• Forming and bending: Forming and bending processes are commonly used to get spacial
components out of copper sheets. Areas which have a high deformation ratio are further
strengthened due to plastic hardening.
• Welding and soldering: Manufacturing operations such as welding and soldering can
be used to get a non-removable metallurgically-bonded connection of dierent stamping-
bending parts.
Instead of assuming an isotropic and location-independent material behaviour throughout the
electrical part, the history of all relevant processes needs to be considered.
1.2 Geometrical inuences

Aside of the manufacturing processes there are several geometrical inuences on the fatigue
properties caused by the load situation. This could be either by a modal resonance of the whole
assembly or a forced displacement due to connected parts. In any case the notch stress situation
inuences the impact of cyclic stresses and the resulting damage accumulation. Geometrical
eects due to a specic load conguration can be handled by dierent concepts:
• Relative stress gradient: The relative stress gradient χ∗ is a parameter to estimate a
supporting eect of the structure. [2]
• High stress volume: Another concept to consider the geometrical eect of notch stresses
is the high stress volume according to a specic value e.g. V90 . The assumption refers to
the statistical probability of defects like pores and cavities in a material volume.
For conventional construction materials like steel and aluminium alloys these two concepts can be
used to describe the transferability of the geometrical inuencing factors. In the FKM-Guideline
[1] it is recommended to consider a supportive factor according to Stieler which is described
by the relative stress gradient χ∗ . Another transferability option is the material mechanics
supportive factor which includes a statistic supportive factor based on the high stress notch area
A90 . Wagener tries to nd a transferability concept for high-copper alloys based on the high
stress volume for standard and small specimens like solder connections in electronics. [3]
2 Test rig setup

To receive valid information on the fatigue strength a test setup is proposed which allows to
stress the components by a constant deection.
Figure 1: Test rig setup for resonance specic S-N curves and test results
The moment of crack initiation can be identied by a drop of the stiness which inuences the
natural frequency of the test specimen. A variation of the load levels and a statistical evaluation
114
allows the derivation of component specic S-N curves. The number of load cycles within the
fatigue relevant area reaches high values up to N = 108 load cycles, which is to be expected of
a cubic face-centered metal such as copper.
Figure 1 shows the test rig setup on an electrodynamic vibration table and the excitation
and response sensor on the specimen. Each specimen is tested until mechanical failure (drop of
stiness) or until a specic number of cycles (run through). With this information a displacement-
controlled S-N curve based on the pearl string method can be derived.
3 Simulations
To examine the transferability of fatigue strength values, the detailed load and stress situation
of the component test as well as the relevant manufacturing processes have to be calculated.
3.1 Load cases

Due to the small size of the specimen and the restricted access to the crack initiation area,
neither strain gauge nor measurement peaks can be applied. To get the local stress and strain
values in the notch, an FE-based calculation will transfer the applied load to the local stress
values. According to the notch stress concept for lifetime estimation, the transfer function will
be linear. This allows to calculate a notch stress S-N curve based on the vibration dwell testing
with constant deection.

It can be shown that the manufacturing processes such as rolling, stamping and forming inu-
ence both the geometrical properties and the local strength. The simulation of the two-step
manufacturing process forming for the treated part shows two relevant areas with high forming
ratios. This results in local cold work hardening and residual stresses. To validate the process
simulation a hardness measurement on the specimen was made along the inner surface of the
bending area. A comparison between the local hardness and the local yield strength shows good
correlation for the process simulation and the real manufactured part.
The next step would be to consider the local material properties for cyclic loads and fatigue
as the failure criterion. Thus, the local distribution of the static strength due to cord work
hardening and residual stresses can be used.
4 Transferability of fatigue strength

The inuence of the real notch geometry can be shown by a comparison of the calculated and
tested crack initiation area. Using a measured 3D-surface geometry of the specimen and a thus
optimized FE-Mesh, the high stress area ts much better with the tested damage pattern.

The manufacturing processes are considered by using either the measured local hardness or
the local yield strength according to a forming process simulation. Even if it was possible to
get a good prediction of the local hardness and yield strength, the fatigue strength cannot be
predicted clearly. The transferability of fatigue parameters is only possible for similar products
and processes, depending on the manufacturing history and the load during lifetime.
115
4.2 Geometrical inuences

The geometrical inuences can be evaluated by the relative stress gradient χ∗ or the high stress
volume V90 . Aside from the diculty to separate the basic eects, the results of this research
show no obvious correlation between notch parameters and fatigue strength for high purity
copper components. Independent of the considered notch parameters the fatigue strength scatters
between a specic notch stress value of approximately σW,a = 126 M P a for N = 3 · 107 load
cycles for dierent specimens.
Figure 2: Application of the common transferability concepts for metals on the test results
(black: milled specimen, blue: stamping-bending specimen)
Whereas typical components based on steel and aluminium alloys have a well known depen-
dency of the relative stress gradient or the high stress volume, the tested specimen in this research
show no signicant dependency. For the high-purity copper which is the focus of this research
no transferability of the geometrical notch eects can be detected according to conventional
methods. This is in accordance to the results of Wagener. [3]
5 Outlook
In this research it is shown that the manufacturing process has a signicant inuence both on
the load distribution and on the fatigue strength. Due to locally dierent material properties the
failure criterion can only be predicted properly based on the process simulation. Geometrical
properties for the notch stress concepts have a minor inuence on the fatigue strength and
thus the transferability. Whenever the manufacturing process has a major impact on the cold
work hardening or a thermal process leads to annealing or recrystallization the geometrical
transferability can be neglected.
Aside from these high-frequency vibrational loads, transient and stationary temperature loads
need to be considered. According to the behaviour of the static strength of the copper material,
a signicant inuence of the ambient temperature on the cyclic strength is to be expected.
References
[1] Forschungskuratorium Maschinenbau (FKM). FKM-Richtlinie - Rechnerischer Festigkeit-
snachweis für Maschinenbauteile. 6. überarbeitete Ausgabe (2012).
[2] Erwin Haibach. Betriebsfestigkeit - Verfahren und Daten zur Bauteilberechnung. Wiesbaden:
Springer, 2006.
[3] R. Wagener, M. Scurria, T. Kimpel, M. Haouel, H. Kaufmann, and T. Melz. Schwingfes-
tigkeitsverhalten von Kupfer-Legierungen für die Anwendung in mobilen Steuergeräten.
METALL 70. Jahrgang 11/2016 (2016), pp. 458 461.
116

Simulation of Fatigue with Cyclic Cohesive Zone

Models
1
Stephan Roth ∗
, Meinhard Kuna1 and Björn Kiefer1
1 TU Bergakademie Freiberg, Institute of Mechanics and Fluid Dynamics, Lampadiusstraße 4, 09596 Freiberg, Germany
Abstract. Cyclic cohesive zone models offer the ability for a unique modeling of fatigue since they comprise
both toughness-based and strength-based fatigue criteria. This way, S–N-curves (Wöhler) and fatigue crack
growth rate curves (Paris) can be predicted with numerical cycle-by-cycle simulations.
In particular, this study addresses the formation and propagation of fatigue cracks in elastic-plastic material.
The capabilities of cyclic cohesive zone models to account for load sequence effects are presented by means
of an illustrative example of a semi-infinite crack under K-controlled loading with a single overload. The
interaction between the damage process zone within the cohesive zone and the plastic zone inside the bulk
material is demonstrated.
1 Cyclic cohesive zone model
The cyclic cohesive zone model (CCZM) applied here to simulate fatigue crack growth (FCG) is briefly
presented in this section [4]. The traction–separation relation of the model is fully described by a cohesive
(state) potential, which is formulated in terms of the normalized effective separation λ = |~s | /s0 and a
fundamental damage variable, 0 ≤ D ≤ ∞, namely
1
Γ λ, D = e − [1 + λF ] exp(1 − λF ) . (1)
F

Here F D represents an auxiliary function which includes the dependency on damage. The coordinates

of the traction vector ~t are calculated by ti /t0 = ∂Γ/∂ si /s0 . Since ~t and ~s are collinear to ensure
rotational equilibrium, the separation law at a particular given damage state can be expressed by

τ λ, D = λF exp(1 − λF ) , (2)
with τ = |~t | /t0 being the normalized effective traction. The parameters cohesive strength t0 and cohesive
length s0 mark the maximum of the traction–separation envelope (TSE)
(
TSE τ λ, D = 1 , ∀λ<1
τ (λ) = , (3)
τ λ, D = λ , ∀λ≥1
which is shown in Fig. 1 (in normalized space), where also some non-linear unloading paths for several
damage states are depicted. The area under the TSE represents the normalized fracture energy density
Γ0 , which in general depends on the shape of the TSE.
Damage initiation and accumulation are controlled by the evolution equation
%
λ
Ḋ = hλ̇i H(λ − λe ) , (4)
D

with the damage exponent % and a state-dependent endurance limit λe D; τ0e , α defined by the normalized
initial endurance limit τ0e and the endurance exponent α. A load scenario illustrating the damage evolution
∗ Corresponding author: stephan.roth@imfd.tu-freiberg.de ORCID iD: 0000-0003-4827-1195
117
is shown in Fig. ??. Further details about the CCZM and its implementation in an ABAQUS user element
subroutine are given in [1, 4].
D= 1 D=1
1
1.5
0.8
2 I II
0.6 2.5
τ
τ
τe III
0.4 3 E
0.2 4 D→∞
5 D→ ∞
0
0 2 4 6 8 λe
λ λ
Figure 1. Monotonic envelope and non-linear un- Figure 2. Initial loading (I), unloading (II), and
loading paths for several damage states reloading (III) paths of the CCZM (en-
D={1, 1.1, 1.5, 2, 2.5, 3, 4, 5, 6} (taken durance limit E); dashed: monotonic en-
from [4]) velope (taken from [4])
2 Application of the CCZM to problems of fatigue
The CCZM presented above has been applied to several problems of fatigue. Since the cohesive model
comprises both strength and toughness type characteristics represented by, e.g., the material parameters
t0 and Γ0 , uniaxial fatigue tests and FCG have been simulated and analyzed using the CCZM, see [2, 4].
It turns out that the particular constitution of the cohesive model—consisting of a potential-based TSE
and an independent damage evolution equation with a damage-dependent endurance limit—additionally
allows to describe the size-dependent macroscopic fatigue failure mode, i.e. a unique modeling of fatigue
where the pure failure modes “uniform debonding” and “fatigue crack growth” represent the material
response for negligible and infinite characteristic length scales, respectively. Consequently, corresponding
limit curves separating fatigue from endurable behavior and static failure in a Kitagawa-Takahashi-
Diagram (KTD) or a Cyclic Failure Assessment Diagram (CFAD) can be predicted, see Fig. ?? and ??.
Note that these limit curves depend on the parameters of the elastic bulk material and the parameters
of the CCZM, compare [3].
K0 static failure
1 static failure
τ0e
FATIGUE
K∞
τ∞
FA
0
Kth
TI
endurable
GU
E
endurable
0.1 0
0.01 0.1 1 10 100 0 τ0e 1
L/Lref τ∞
Figure 3. KTD with failure limit curves predicted Figure 4. CFAD with failure limit curves predicted
by the CCZM (taken from [3]) by the CCZM (taken from [3])
3 Combination of CCZM with elastic-plastic bulk material
In all previous studies, the CCZM has been utilized in combination with linear-elastic bulk material,
using the basic assumption that all inelastic processes are covered by the cohesive behavior. Now, the
interplay of the cohesive model with surrounding elastic-plastic material is analyzed. The boundary layer
118
ur/ϕ (KI )
out
r
x2
x2
x1 , ∆a
ϕ x1 , r, ∆a
a) Lfine b) h
Figure 5. a) Crack model with KI -controlled displacement boundary conditions; b) detailed mesh within the finely
meshed region 0 ≤ x1 ≤ Lfine (taken from [4])
model of a semi-infinite crack under plane strain conditions subjected to Mode-I loading is considered,
see Fig. ??. More details can be found in [4]. In the vicinity of the crack tip, J2 -plasticity with non-linear
isotropic hardening according to the power-law
" #n
σy σy E
= + εeq , (5)
σy0 σy0 σy0
with yield stress σy , initial yield stress σy0 , Young’s modulus E, equivalent plastic strain εeq , and
hardening exponent n is applied. The chosen geometric and constitutive parameters are summarized in
Tab. ??:
Table 1. Parameters of the cohesive crack model: geometry, material, loading

rout /s0 Lfine /s0 h/s0 E/σy0 ν σy0 n t0 /σy0 s0 Γ0 % τ0e α ∆K/K0 R
108 20 − 104 0.5 340 0.3 1 0.1 3.4 1 2.7 1 0.2 2 0.6 0
For a given load ratio R = KImin /KImax = 0, FCG rate curves were computed based on simula-
tions with
p a maximum stress intensity factor varying in a range of 0.15 ≤ KImax /K0 ≤ 1.2 with
K0 = E t0 s0 Γ0 / [1 − ν ] and ∆K = KI . In order to demonstrate the influence of plasticity, a
2 max
reference FCG rate curve for a ratio t0 /σy0 = 0.5 representing linear elasticity was generated. The nor-
malized FCG rate curves are depicted in Fig. ??. For the given set of material parameters, both curves
coincide in the near-threshold and Paris regions up to ∆K ≈ 0.5K0 . With further increase of the load
level, considerable cyclic plastic deformation and thus dissipation occurs leading to a deviation of the
FCG rate curves. While the elastic curve approaches static failure at ∆K = K0 , FCG also occurs at
higher load levels in case of plastic bulk material.
Significant differences also appear when the FCG behavior is observed after single overloads. Here, as
illustrated in Fig. ??, an overload of ∆K = 0.9K0 is applied after steady-state FCG has been established,
i.e. after 50 load cycles (N = 50), at a basic load level of ∆K = 0.6K0 . At steady-state, the FCG rates for
elastic and elastic-plastic material behavior are nearly the same, compare Fig. ??. Also, the magnitudes
of the crack length jumps in consequence of the overload do not differ. After the single overload, when the
basic cyclic load level ∆K = 0.6K0 is again applied, the elastic model quickly returns to the former FCG
rate, thus predicting an acceleration of FCG. In contrast, the plastic model shows a pronounced crack
retardation before the steady-state is reached again. To explain this effect, the distributions of equivalent
plastic strain (in the bulk material) and damage (in the cohesive zone) is visualized in Fig. ??–?? at
i) N = 1 (primary plastic zone, PPZ), ii) N = 50 (steady-state), iii) N = 51 (overload), iv) N = 85
(crack retardation), and v) N = 101 (return to steady-state). After leaving the primary plastic zone, the
steady-state is characterized by a constant FCG rate accompanied by a growth of the cyclic plastic zone
(CPZ) around the fatigue crack tip. The overload causes a sudden crack jump and an enlargement of the
plastic zone. This region has to be crossed subsequently by the FCG crack inducing the observed crack
retardation. Once the region affected by the overload is left, the FCG crack returns to the steady-state.
119
∆J/ [t0 s0 Γ0 ]
0.0 0.1 0.5 1.0
600
1E+2 elastic 500

plastic
400
∆a/s0
d ∆a/s0 /dN

300
1E+0
200 elastic
elastic, overload
100 elastic-plastic
1E−2 elastic-plastic, overload
0
0 10 20 30 40 50 60 70 80 90
0.1 0.5 1.0
N
∆K/K0
Figure 6. FCG rate curves for elastic and plastic Figure 7. Fatigue crack extension ∆a(N ) with and
(t0 /σy0 = 3.4) bulk material without single overload at N = 51
Figure 8. PPZ: N = 1 Figure 9. Steady-state: N = 50 Figure 10. Overload: N = 51
Figure 11. Crack retardation: N = 85 Figure 12. Return to steady-state: N = 101
4 Summary
In this study, the combination of a plastic bulk material and a CCZM to simulate FCG was investigated.
For a representative ratio t0 /σy0 , the FCG rate curve was computed and compared with the elastic
curve. The typical behavior of crack propagation in the near-threshold and Paris-regions is reflected.
As expected, due to plastic dissipation, static failure occurs at considerably higher load levels. The
application of a single overload causes crack retardation, which also matches experimental observations.
References
[1] S. Roth. “Entwicklung und Implementierung zyklischer Kohäsivzonenmodelle zur Simulation von
Werkstoffermüdung”. Dissertation. TU Bergakademie Freiberg, 2016.
[2] S. Roth and B. Kiefer. Simulation of fatigue crack growth with a cyclic cohesive zone model using a
phase field description. In: DVM-Bericht 251. DVM. 2019, pp. 159–168.
[3] S. Roth and M. Kuna. Prediction of size-dependent fatigue failure modes by means of a cyclic cohesive
zone model. Int. J. Fatigue 100.1 (2017), pp. 58–67.
[4] S. Roth, G. Hütter, and M. Kuna. Simulation of fatigue crack growth with a cyclic cohesive zone
model. Int. J. Fract. 188.1 (2014), pp. 23–45.
120

Efficient phase-field modelling of fatigue in ductile

materials
1
Martha Seiler ∗
, Thomas Linse1 , Peter Hantschke2 and Markus Kästner1
1 Chair of Computational and Experimental Solid Mechanics, TU Dresden, 01062 Dresden, Germany
2 Structural Durability Group, TU Dresden, 01062 Dresden, Germany
Abstract. Fatigue loads come along with high numbers of load cycles which makes the simulation of fatigue
fracture computationally very demanding. We present a numerically efficient method which avoids resolving
the loading path and instead requires only one increment per load cycle at most. We combine the phase-field
method for fracture with a classic fatigue concept. A local lifetime variable is obtained, which we use to
modify the material resistance incrementally in order to consider the progressive weakening of the material.
Focussing on ductile material behaviour, we make use of an elasto-plastic revaluation technique instead of a
full plastic material model. The model allows to describe fatigue crack initiation as well as propagation and
reproduces Paris behaviour.
Introduction
Fatigue fracture is one of the most common causes of failure in structures, while still being insufficiently
predictable. Fatigue life is divided in a crack initiation and a crack propagation stage. Stable cyclic
crack propagation can be described by the Paris-Erdogan law [5]: It links the cyclic crack propagation
rate da/dN to the range of stress intensity factor ∆K by the power law da/dN = C∆K m with material
parameters C and m. For the numerical simulation of crack propagation, the phase-field method is a
very promising approach, being able to model both crack initiation and propagation in a straightforward
manner [3]. Carrara et al. [1] applied the method to fatigue fracture in brittle materials, decreasing the
crack resistance due to repetitive loading. Analogously, we introduce a model for ductile materials such
as metals. It is also based on the reduction of the fracture toughness, but considers that fatigue crack
propagation in ductile materials is caused by cyclic plasticity.
1 Model framework
The model is based on the phase-field formulation for brittle fracture introduced by Miehe et al. [3]. In
order to describe the crack topology, an additional phase-field variable d ∈ [0, 1] is introduced, smoothly
bridging the fully intact (d = 0) and fully damaged (d = 1) state. The degradation function g(d) =
(1 − d)2 couples the mechanical field u and the phase-field d, degrading the stress σ = g(d) ∂ψ e /∂ε.
Approximating the crack geometry by regularizing the phase-field with the length scale parameter `, the
regularized energy functional can be written as
Z Z
1
Π` = g(d) ψ e (ε) dV + α(D) Gc (d2 + `2 |∇d|2 ) dV. (1)
Ω Ω 2`
The standard phase-field formulation is thereby extended to fatigue by introducing a scalar reduction
function α(D) ∈ [α0 , 1] with 0 < α0 < 1. It reduces the fracture toughness Gc locally in order to model
the progressive material weakening. This reduction function depends on a lifetime variable D. For D = 0
a material point has experienced no fatigue loads at all, while D = 1 means it has undergone all load
cycles it can possibly bear before loosing its integrity.
∗ Corresponding author: martha.seiler@tu-dresden.de
121
On this basis the function

α(D) = (1 − α0 )(1 − D)ξ + α0 (2)
with the parameters α0 and ξ is proposed. The governing equations for the coupled problem
~0 = ∇ · σ α(D) d − ∇α(D) · `2 ∇d − α(D) `2 ∆d = (1 − d) ψ e (ε)2`/Gc (3)
are complemented with boundary conditions n · σ = t̃, u = ũ and n · ∇d = 0. We set the irreversibility
condition d˙ ≥ 0. For a discussion of the influence of irreversibility see Linse et al. [2].
The lifetime variable D is determined with the notch strain concept [6], which is in its original sense a
common scheme for service life prediction of components. Stresses and strains derived from the solution of
the displacement field u are revaluated to virtual elasto-plastic values following Neuber’s rule[4]. Hence,
no plastic material model is needed, which significantly reduces computational time. With the revaluated
stress and strains, the damage parameter PSWT [7] is determined, which is associated with the area inside
the stress-strain hysteresis curve. It only depends on the stress and strain amplitudes σa and εa and the
mean stress σm . Hence, for alternating loads one elastic simulation per load cycle at amplitude load is
sufficient in order to determine its damaging effect. With the damage parameter the lifetime contribution
∆Di of a load cycle can be determined from strain Wöhler curves, which are derived from standardised
cyclic experiments.
The method is tested with a compact tension (CT) test. The specimen is loaded with a constant force
amplitude of F̃ = 4.9 N and no mean load. The model parameters are chosen to E = 210 GPa, ν = 0.3
and Gc = 2.7 N/mm, α0 = 0.05 and ξ = 1. The set-up as well as evolution of the phase-field and the
lifetime variable are displayed in Fig. 1. Starting from the notch, a zone of D = 1 forms progressively.
This area can be associated with the plastic zone, since D is computed from the damage parameter PSWT ,
which is again associated with the area inside a stress-strain hysteresis – a measure for dissipation.
N 1165 1202 1240

1.0
F̃ 0.325
0.275 d 0.5
a
0.0
0.6
1.0
T =1
F̃
D 0.5
0.125 0.2 0.8
0.0
Figure 1. Compact tension test. All dimensions in mm. Development of phase-field variable d and lifetime variable
D for load cycles N .
The resulting crack propagation curves for varying F̃ are displayed in Fig. 2, showing a crack initiation
as well as a propagation stage. The Paris plot in Fig. 3a shows that the method can repoduce Paris
behaviour: The Paris parameters C and m, fitted to the stable crack propagation stage, are independent
of the load. Being material parameters, they are influenced by material properties instead, as Fig. 3b
shows by varying the hardening modulus K 0 of the cyclic stress-strain curve.
122
F̃ = 0.0042 kN
F̃ = 0.0046 kN
0.3
F̃ = 0.0049 kN
F̃ = 0.0056 kN
0.2
a
0.1
0.0
0 2000 4000 6000 8000 10000 12000
N
Figure 2. Cyclic crack propagation in CT specimen. N - cycle number, a - crack length, F̃ - force amplitude.
10−1
F̃ = 0.0042 kN K0 = 1.500 GPa
0
10 K0 = 1.800 GPa
F̃ = 0.0046 kN
F̃ = 0.0049 kN K0 = 2.115 GPa
10−1 10−2
∆a/∆N /mm
∆a/∆N /mm
F̃ = 0.0056 kN K0 = 2.500 GPa
K0 = 3.000 GPa
10−2
10−3
10−3
10−4 10−4
3 × 10−1 4 × 10−1 6 × 10−1 3 × 10−1 4 × 10−1 6 × 10−1

√ √
∆K/GPa mm (a) ∆K/GPa mm (b)
Figure 3. Paris plot of cyclic crack propagation in CT specimen. ∆a/∆N - crack propagation rate, ∆K - stress
intensity factor range, F̃ - force amplitude, K 0 - material parameter of the cyclic stress-strain curve.
Acknowledgments
This work was supported by the German Research Foundation DFG within the Priority Programme 2013
Targeted Use of Forming Induced Residual Stresses in Metal Components grant number KA 3309/7-1.
References
[1] P. Carrara, M. Ambati, R. Alessi, and L. De Lorenzis. A novel framework to model the fatigue
behavior of brittle materials based on a variational phase-field approach (2018). arXiv: 1811.02244.
[2] T. Linse, P. Hennig, M. Kästner, and R. de Borst. A convergence study of phase-field models for
brittle fracture. Engineering Fracture Mechanics 184.Supplement C (2017), pp. 307 –318.
[3] C. Miehe, F. Welschinger, and M. Hofacker. Thermodynamically consistent phase-field models of
fracture: Variational principles and multi-field FE implementations. International Journal for Nu-
merical Methods in Engineering 83.10 (2010), pp. 1273–1311.
[4] H. Neuber. Theory of Stress Concentration for Shear-Strained Prismatical Bodies With Arbitrary
Nonlinear Stress-Strain Law. Journal of Applied Mechanics 28.4 (1961), pp. 544 –550.
[5] P. Paris and F. Erdogan. A Critical Analysis of Crack Propagation Laws. Journal of Basic Engi-
neering 85.4 (1963), p. 528.
[6] T Seeger. Grundlagen für Betriebsfestigkeitsnachweise. Fundamentals for Service Fatigue-Strength
Assessments),” Stahlbau Handbuch (Handbook of Structural Engineering), Stahlbau-Verlags-gesellschaft,
Cologne 1 (1996), pp. 5–123.
[7] K. N. Smith, T. Topper, and P. Watson. A stress-strain function for the fatigue of metals. Journal
of Materials 5 (1970), pp. 767–778.
123
124

High-order cycle jump integration of a concrete fatigue

damage model
1
Thomas Titscher ∗
, Jörg F. Unger1 and Javier Oliver2
1 Federal Institute for Materials Research and Testing (BAM)
2 International Center for Numerical Methods in Engineering (CIMNE), Technical Univ. of Catalonia (BarcelonaTech)
Abstract. Fatigue models that accurately resolve the complex three-dimensional failure mechanisms of concrete
are numerically expensive. Especially the calibration of fatigue parameters to existing Wöhler lines requires
solving for thousands or millions of cycles and a naive cycle-by-cycle integration is not feasible. The proposed
adaptive cycle jump methods provide a remedy to this challenge. They greatly reduce the numerical effort of
fatigue simulations and provide the basis for a development of those models.
1 Introduction
Optimizing and maximizing the structures performance under fatigue load has the potential of saving
resources and energy. Especially for concrete, this requires precise models to handle the complexity of
the material. A common approach for the lifetime estimation are Wöhler lines that relate an external
cyclic load to the number of loading cycles a test specimen endures. An elastic simulation of the structure
identifies the highest stresses and the corresponding lifetime estimation is read from the Wöhler lines.
It is challenging to relate the test specimen results to a structural level, because they are influenced
by numerous factors, e.g. loading frequency, environmental conditions, size effects, stress redistributions
and multiphysics effects. This results in enormous experimental efforts to create Wöhler curves for all
combinations of effects.
A more general approach that is commonly used in low cycle fatigue[3] is the cycle-by-cycle integration.
A single load cycle is divided into several load steps and the structures equilibrium state is computed for
each of those cycles. That approach does not rely on experimental Wöhler lines as an input to the material
law, but has two main challenges that will be addressed in the following work. First, sophisticated fatigue
models are needed that accumulate a failure defining variable in each cycle. Second, the high numerical
effort has to be reduced. Compared to the Wöhler approach, however, this has additional benefits. It
allows modeling of stress redistributions over the lifetime, it is independent of the load shape and it can
lead to deeper understanding of the processes that lead to fatigue failure.
The changing state of a material is described by the change of its history variables κ. For cyclic loading,
this change ∆κ is small in a single cycle. But over the lifetime of the material, it changes significantly
and may lead to material failure. This scale separation gives rise to the vivid idea of cycle jump where
the small change over one cycle is extrapolated.
This idea can be generalized to an explicit ordinary differential equation that relates the global change of
the history variables over the lifetime with the local change over one cycle. It reads
dκ
= ∆κ(κ, N ), (1)
dN
where N denotes the cycle number. Its left-hand side can be approximated with suitable, well-known
integration schemes and results in cycle jump methods. The evaluation of the right-hand side typically
involves the solution of a finite element model that is capable of reproducing the material behavior. Both
problems are now disconnected and can be evaluated and assessed separately.
∗ Corresponding author: thomas.titscher@bam.de ORCID iD: 0000-0001-9470-281X
125
2 Adaptive cycle jump methods
The first order Euler forward discretization of the left-hand side in Eq. (1) corresponds to the direct cycle
jump method[1] and is illustrated in Fig. 1. Higher-order methods like the midpoint rule[6] or a fourth
order Runge-Kutta method[2] provide more accurate approximations at the cost of more evaluations of
the right-hand side.
cyclic external load

dependent variable
history variable κ + ∆ N ∆κ N
κN +∆N = κN
∆κN
tN tN + τ tN +∆N tN +∆N + τ
time t
Figure 1. Direct cycle jump method based on an Euler forward approximation of the history variables.
In the first part of the lifetime, the structures state is dominated by static effects and changes rapidly.
This requires small jump lengths or even a direct numerical simulation (DNS) of those cycles to minimize
the error of the cycle jump approximation. In the next region of steady material degradation, the jump
length can be increased. Material failure then happens rapidly and, again, requires small step sizes. So
efficient implementations have to provide a measure to adapt the jump length.
For this purpose, the adaptive Runge-Kutta schemes Bogacki-Shampine(3(2)) and Dormand-Prince(5(4))
are explored. They provide an algorithm to adaptively determine the step size (jump length). In the case
of Dormand-Prince(5(4)), both a fourth and a fifth order approximation is provided in each step. The
difference of both approximations is compared to a prescribed tolerance δ to a) decide to accept or decline
the fifth order step and b) determine the next jump length.
3 Fatigue extension of the implicit gradient-enhanced damage model
In this work, the material behavior is described by the implicit gradient-enhanced damage model by
Peerlings et al.[5]. It consists of the momentum balance equation
∇ · σ = ∇ · [(1 − ω)C : ε] = 0 (2)
with the second order strain and stress tensors ε and σ, respectively, the fourth order elasticity tensor C
and the isotropic damage variable ω ∈ [0, 1). Damage is driven by the scalar strain-like history variable
κ that only grows on the damage surface
g(κ) = ε̄ − κ = 0, (3)
which is described by the Karush–Kuhn–Tucker conditions
κ̇ ≥ 0, g ≤ 0, g κ̇ = 0. (4)
In Eq. (3), ε̄ are the nonlocal strains that are related to a local strain norm εeq via the screened Poisson
equation
ε̄ − l2 ∇2 ε̄ = εeq (ε). (5)
It limits the curvature of the nonlocal strains ε̄ depending on the length parameter l and, thus, regularizes
the model.
126
Equation (4) for static material failure is extended to fatigue failure by allowing the growth of κ below
the damage surface[4]
(
ε̇eq if g = 0
κ̇ = (6)
˙ + f (κ, ε̄) if g < 0.
hε̄i
The Macaulay brackets denoted by h i+ correspond to the ramp function and prohibit a damage accu-
mulation in unloading conditions and the fatigue function f controls the actual fatigue behavior.
4 Results
The model is exposed to cyclic loading to investigate the fatigue properties. Both displacement and
force controlled simulations are performed in a one-dimensional uniaxial setting to evaluate the fatigue
function f . Basic expected fatigue behavior like the lifetime decrease for increased loading amplitudes
is recovered. A compact tension specimen under cyclic loading shows the stress redistribution over the
specimen’s lifetime. The material at the notch tip cannot withstand the initial stress concentration.
Damage occurs and deflects the stresses to a slightly larger region. This process continues and widens the
notch tip.
Both experiments are also integrated with the cycle jump methods, which drastically reduces the number
of cycle evaluations (the main numerical effort) to about 1-2% of the direct simulation. The accuracy of
this approximation varies both in terms of resolving a reference solution after a given number of cycles
and in resolving the failure point. The first case can be efficiently solved with cycle jump methods and
a fixed jump length. Here, second order method clearly outperform first order methods. The benefit of
the forth order methods, however, is ambiguous. They require a longer jump length to reach comparable
computational efficiency which can lead to significant deviations from the solution in the early fatigue
life. The second case of resolving material failure practically requires adaptive cycle jump methods as
the final number of cycles is unknown. Here, the accuracy of the commonly used Dormand-Prince(5(4))
method is similar to the numerically cheaper Bogacki-Shampine(3(2)) method.
The Bogacki-Shampine method is then used to simulate Wöhler lines of three-dimensional cylinder spec-
imen under cyclic loading. Different combinations of fatigue parameters either shift the resulting Wöhler
line or modify its slope and allow the calibration to experimental data.
References
[1] D Cojocaru and AM Karlsson. A simple numerical method of cycle jumps for cyclically loaded
structures. International Journal of fatigue 28.12 (2006), pp. 1677–1689.
[2] VM Kindrachuk and JF Unger. A Fourier transformation-based temporal integration scheme for
viscoplastic solids subjected to fatigue deterioration. International Journal of Fatigue 100 (2017),
pp. 215–228.
[3] M Kuna and S Roth. General remarks on cyclic cohesive zone models. International Journal of
Fracture 196.1-2 (2015), pp. 147–167.
[4] JJ Marigo. Modelling of brittle and fatigue damage for elastic material by growth of microvoids.
Engineering Fracture Mechanics 21.4 (1985), pp. 861–874.
[5] RHJ Peerlings, R De Borst, WAM Brekelmans, and JHP De Vree. Gradient enhanced damage for
quasi-brittle materials. International Journal for Numerical Methods in Engineering 39.19 (1996),
pp. 3391–3403.
[6] RHJ Peerlings, WAM Brekelmans, R de Borst, and MGD Geers. Gradient-enhanced damage mod-
elling of high-cycle fatigue. International Journal for Numerical Methods in Engineering 49.12 (2000),
pp. 1547–1569.
127
128
MS 07: Higher Order Numerical Methods in Space and Time

Mixed nite element formulations and energy-momentum

time integrators for thermo-mechanically coupled ber-
reinforced continua
1
Dietzsch, J. , Groÿ, M.
2 and Flessing, L.
3
1 Technische Universität Chemnitz, Professorship of applied mechanics and dynamics, Reichenhainer Straÿe 70, D-09126 Chemnitz,
julian.dietzsch@mb.tu-chemnitz.de
2 michael.gross@mb.tu-chemnitz.de
3 tmd@mb.tu-chemnitz.de
Abstract. Fiber-reinforced materials in lightweight structures and their accurate dynamic simulation play an
even increasingly signicant role today. These materials are used for example in aircraft, automobiles and wind
turbines. The low density and the high modulus of elasticity play a major role, but also the thermal properties
should not be neglected. First of all, the thermal expansion of the matrix part and the ability to conduct the
heat in a directional way with the bers. For these materials, volumetric locking eects of an incompressible
matrix material as well as locking eects due to sti bers can occur. On the one hand, there are combinations
of well known mixed elements with an independent approximation of the volumetric dilation of the deformation
gradient and an independent approximation of the right Cauchy green tensor for the anisotropic part of the
strain energy function to reduce these eects. On the other hand, we have developed mixed elements where
elds for the fourth and fth invariants are added, as well as a version with the corresponding tensor elds.
For long-term simulations it is necessary to use higher order time integrators to perform an accurate dynamic
simulation. Galerkin-based time integrators oer a good option for this application. To eliminate a huge energy
error these have to be extended to an energy-momentum time integration scheme. It is logical to combine
these methods and thus combine the advantages of these methods. We formulate the mixed elements using
Hu-Washizu functionals and combine this with the mixed principle of virtual power. By adding an thermo-
mechanical coupling part in the strain energy and introducing the fourier heat conduction we obtain a thermo-
mechanical formulation for the dierent mixed elements and a higher order Galerkin-based time integrator.
Dirichlet boundary conditions in the form of Lagrange multiplier methods as well as Neumann boundary
conditions in the mechanical and thermal context are also provided. In addition, we extend the continuum
so that we can model dierent ber families and the directional heat conduction of the bers. As numerical
examples serves cook's cantilever beam as well as a rotating heat pipe. We primarily analyze the spatial
and time convergence, the conservation properties as well as the eect of the heat conduction of the bers.
1 Introduction
As continuum model, we consider an anisotropic material with nF ber directions ai0 moving in the
Euclidean space Rndim with the constant ambient temperature Θ∞ . The strain energy function of the
material with a thermo-viscoelastic matrix and thermoelastic bers is given by
nF
X
Ψ(C, C v , Θ) = ΨM (C, C v , Θ) + ΨFi (C, Θ, M i ), (1)
i=1
which is split into a single matrix part ΨM and multiple nF ber parts ΨFi . Here, M i = ai0 ⊗ai0 dene the
structural tensor, C = F T F dene the right Cauchy-Green tensor, C v dene the viscous rightpCauchy-
Green tensor and Θ dene the absolute temperature. With the volumen dilatation J = det[F ] = det[C],
we assume the specic dependencies:
cap coup
ΨM (C, C v , Θ) = Ψiso vol vis
M (C, cof[C], J) + ΨM (J) + ΨM (Θ) + ΨM (Θ, J) + ΨM (Λ) (2)
ΨFi (C, Θ, M i ) = Ψela
Fi (C, cof[C], J, M i ) + Ψcap
Fi (Θ) + Ψcoup
Fi (Θ, C, M i ) (3)
The elastic part of the matrix function ΨM is split into an isochoric part Ψiso
M and a volumetric part
129
1 = 0.1e6 Vdev = 0.1e4 kF1 = 0.1

A
2 = 0.1e6 Vvol = 0.2e4 βF1 = 1e − 10
y p̂
0.16
3 = 1.8e6 6 = 10e6 c0F1 = 1500
4 = 100e6 7 = 4 c1F1 = 0.003
5 = 4 8 = 4 kF2 = 50
vis
1 = 0.1e5 9 = 1 βF2 = 1e − 10
0.44
0.44 vis
2 = 0.1e5 10 = 0.1e6 c0F2 = 3000
B
vis
3 = 1.8e5 kM = 0.1 c1F2 = 0.006
vis
4 = 100e5 βM = 1e − 10 Θ∞ = 300
x vis
5 =4 c0M = 1500 p̂ = 1.5e6
0.48
0.10
z
ρ0 = 1000 c1M = 0.003
hn = 0.005 T = 1.5 T OL = 1e − 4
Figure 1. Geometry, conguration, simulation parameters and ber directions (a10 )T = [1 1 1] (blue) and (a20 )T =
[1 1 0] (green) of cooks cantilever beam.
cap
M . We subdivided the thermo-elastic free energy of the matrix into a heat capacity part ΨM and the
Ψvol
part of the thermo-mechanical coupling eect, which takes the form
∂Ψvol
M (J)
Ψcoup
M (Θ, J) = −2ndim βM (Θ − Θ∞ )J , (4)
∂J
where βM ist the coecient of linear thermal expansion for the matrix. The function Ψvis
M is the viscoelastic
free energy function of the matrix material, with Λ = CC −1 v . The thermal part of the ber free energy
is separated in the same manner. We consider a heat capacity function Ψcap Fi and the function of the
thermo-mechanical coupling
q ∂Ψela (I i , . . . )
Fi 4
Ψcoup i
Fi (Θ, C, M i ) = −2βFi (Θ − Θ∞ ) I4 , (5)
∂I4i
with the coecients of linear thermal expansion βFi and the fourth invariant I4i (C, M i ) = tr[CM i ].
As shown in Reference [1], using the CoCoA element prevents looking in an ecient manner. This nite
element formulation is based on a Hu-Washizu functional. The complete functional of the internal energy
is given by
Z Z Z Z
Πint = ΨdV + η (Θ − Θ̃)dV + B : (cof[C(q)] − H)dV + ˜
p̃ (J(q) − J)dV
B0 B0 B0 B0
Z Z Z
1
+ S A : (C(q) − C A )dV + B A : (cof[C(q)] − H A )dV + p̃A (J(q) − JÃ )dV (6)
B0 2 B0 B0
With independent variables for the assumed temperature eld Θ̃, for the volumetric dilatation J˜ (see
Reference [5]), H for the cofactor of C (see Reference [3]) , for the anisotropic parts ΨFi of the material
formulation C A (see Reference [4]) and H A and JÃ (see Reference [1]). Notable Lagrange multipliers
are p̃ wich plays the role of the hydrostatic pressure, S A as anisotropic part of the stress tensor and the
entropy density eld η . Thus the dependencies in the strain energy function change as follows:
˜ vol ˜ cap coup ˜
ΨM = Ψiso
M (C, H, J) + ΨM (J) + ΨM (Θ) + ΨM (Θ, J) + ΨM (Λ)
vis
(7)
ΨF = Ψela (C A , H A , JÃ , M i ) + Ψcap (Θ) + Ψcoup (Θ, C A , M i )
i Fi Fi Fi (8)
In order to obtain the weak form, we use the mixed principle of virtual power, as shown in Reference [2].
The basis is the total energy balance
˜˙ B, Ḣ, S A , Ċ A , B A , Ḣ A , p̃A , J˙A ) = 0

Ṫ (q̇, v̇, ṗ) + Π̇ext (q̇, λ, Ċ v , Θ̃) + Π̇int (q̇, Ċ v , Θ̇, η̇, Θ̃, p̃, J, (9)
where the time derivative of the kinetic energy

Z Z Z
Ṫ (q̇, v̇, ṗ) = (ρ0 v − p) · v̇dV + ṗ · (q̇ − v)dV + p · q̈dV (10)
B0 B0 B0
130
is dened by the velocity v , the linear momentum p and the mass density ρ0 . As external power functional,
we assume
Z Z Z !
ext ref Θ̃
Π̇ (q̇, λ, Ċ v , Θ̃) = − t · q̇dA − λ · (q̇ − q̇ )dA + ∇ · QdV
∂B0 ∂B0 B0 Θ
Z Z "n #
Θ̃ int X F
kFi − kM −1
+ D dV + Ċ v : V(C v ) : Ċ v dV Q=− J M i + kJC ∇Θ (11)
B0 Θ B0 i=1
C : Mi
Here, Q denotes the Piola heat ux vector derived from Duhamel's law (see Reference [2]). Where kM
and kFi denotes the material conductivity coecients for matrix and bers. The eld q̇ ref denotes the
time evolution of the Dirichlet boundary displacement vector with the corresponding Lagrange multiplier
λ. And t denotes the traction load for the Neumann boundary condition. The non-negative internal
viscous dissipation Dint and positive-denite viscosity tensor V ist given by

1 Vdev Vdev
int
D = Ċ v : V(C v ) : Ċ v V(C v ) = Vvol − C −1 −1
v ⊗ Cv + Is : C −1 −1
v ⊗ Cv , (12)
4 ndim 4
with the fourth-order symmetric projection tensor Is and the material paramenters Vdev and Vvol which
represent the volumetric and deviatoric viscosity constants, respectively. By variation with respect to the
variables in the argument from Eqn. (9) (as shown in Reference [2]) we obtain the weak forms. Thereby
all quantities are approximated with Lagrangian shape functions in space (see Reference [2]) and time
(see Reference [2]). Each integral that exist are solved with the corresponding Gaussian quadrature rule.
The internal variable C v is determined on the element using the Newton-Raphson method, but not at
each spatial quadrature point by using spatial nite element shape functions. After eliminating p and
η , we condense out the resulting formulation at the element level to a displacement and temperature
formulation (see Reference [3]). Therefore, all mixed elds except q and Θ are discontinuous at the
boundaries of spatial elements.
2 Numerical Example
As numerical example serves the well-known Cook's cantilever beam with a quadratic distribution of an
in-plane load on the Neumann boundary. Two dierent bers are used for this example. On the one side
a ber (F1 ,(a10 )T = [1 1 1]) for mechanical reinforcement of the matrix, but with an equally low thermal
conductivity. On the other side, a ber (F2 ,(a20 )T = [1 1 0]) with a very high thermal conductivity, but
a low stiness. The energy functions are given by
1 2 4 5
Ψiso
M = (tr[C])2 + (tr[cof[C]])2 − 3 ln(J) Ψvol M = (J + J −5 − 2)
2 2 2
vis vis vis vis vis
Ψvis
M = 1 (tr[C])2 + 2 (tr[cof[C]])2 − vis 3 ln(J) +
4
(J 5 + J −5 − 2)
2 2 2
1 1 1 −
Ψela
F1 = 6 (tr[CM 1 ])7 +1
+ (tr[cof[C]M 1 ])8 +1 + det[C] 9
7 + 1 8 + 1 9
10 Θ 1
Ψela
F2 = (tr[CM 2 ] − 1)2 Ψcap 0 1
X = cX (1 − Θ∞ cX )(Θ − Θ∞ − Θ ln ) − c0X c1X (Θ − Θ∞ )2
2 Θ∞ 2
and the prescribed simulation parameters, geometry, conguration and ber directions are shown in Fig. 1.
We use the CoCoA210000 element, with an quadratic approximation of q and a linear approximation of
H . All other quantities are approximated constant. This selection is based on the results of Reference [1],
where this element showed an excellent convergence rate. In the rst step we only use F1 . The result
can be found in part 1 and 2 of Fig. 2. First we see the asymmetric bending caused by the ber as
well as the typical stress curve (tensile and compressive stress, neutral ber). Together with the small
number of elements, the excellent convergence rate of the mixed elements is once again clearly shown.
Regarding the temperature distribution we can see that in some areas, as a result of the dissipation of
the viscous part, the temperature increases. If we now add the second ber, these temperature increases
are distributed according to the ber direction.
131
HCoCo210000 HCoCo210000 HCoCo210000

t = 1.5 t = 1.5 t = 1.5
nF = 1 nF = 1 nF = 2
Θ Θ
σ VM 300.10 300.10
4.0e6
2.7e6
300.05 300.05
1.4e6
5.0e4 300.00 300.00
Figure 2. Deformed conguration Bt and v. Mises equivalent stress σV M and the absolute temperature θ for the
parameters shown in Fig 1 and bers F1 and F2 for t = 1.2. Cook's cantilever beam with nel = 32
spatial nite elements.
3 Conclusions
We were able to show that the extension to an thermo-viscoelastic material formulation, do not aect
the excellent convergence rates of the mixed elements. This saves computational time, which is especially
advantageous with regard to the local iteration of the viscous internal variable. In the next step, we
extend the ber part by a viscous component and we extend this formulation to an energy conserving
time integrator.
Acknowledgments
The authors thank the 'Deutsche Forschungsgesellschaft (DFG)' for nancial support of this work under
the grant GR 3297/4-1 and Matthias Bartelt (GR 3297/2-2) for providing the program code and help
with its implementation.
References
[1] Julian Dietzsch, Michael Groÿ, and Lars Flessing. THERMO-MECHANICAL COUPLING IN FIBER-
REINFORCED CONTINUA: MIXED FINITE ELEMENT FORMULATIONS AND ENERGY-
MOMENTUM TIME INTEGRATION. In: VIII International Conference on Computational Meth-
ods for Coupled Problems in Science and Engineering. Sitges (Barcelona), Spain, June 2019.
[2] Michael Groÿ, Julian Dietzsch, and Matthias Bartelt. Variational-based higher-order accurate en-
ergymomentum schemes for thermo-viscoelastic ber-reinforced continua. Computer Methods in
Applied Mechanics and Engineering 336 (2018), pp. 353 418. issn: 0045-7825. doi: https://doi.
org/10.1016/j.cma.2018.03.019. url: http://www.sciencedirect.com/science/article/
pii/S0045782518301440.
[3] Jörg Schröder, Peter Wriggers, and Daniel Balzani. A new mixed nite element based on dierent ap-
proximations of the minors of deformation tensors. Computer Methods in Applied Mechanics and En-
gineering 200.49 (2011), pp. 3583 3600. issn: 0045-7825. doi: https://doi.org/10.1016/j.cma.
2011.08.009. url: http://www.sciencedirect.com/science/article/pii/S0045782511002647.
[4] Jörg Schröder, Nils Viebahn, Daniel Balzani, and Peter Wriggers. A novel mixed nite element
for nite anisotropic elasticity; the SKA-element Simplied Kinematics for Anisotropy. Computer
Methods in Applied Mechanics and Engineering 310 (2016), pp. 475 494. issn: 0045-7825. doi:
https : / / doi . org / 10 . 1016 / j . cma . 2016 . 06 . 029. url: http : / / www . sciencedirect . com /
science/article/pii/S0045782516306478.
[5] J.C. Simo, R.L. Taylor, and K.S. Pister. Variational and projection methods for the volume constraint
in nite deformation elasto-plasticity. Computer Methods in Applied Mechanics and Engineering 51.1
(1985), pp. 177 208. issn: 0045-7825. doi: https://doi.org/10.1016/0045-7825(85)90033-7.
url: http://www.sciencedirect.com/science/article/pii/0045782585900337.
132

Application of diagonally implicit Runge Kutta methods

to structural dynamics problems
Matthias Grafenhorst∗1 and Stefan Hartmann1
1 Institute of Applied Mechanics, Clausthal University of Technology, Clausthal-Zellerfeld, Germany
Abstract. The most important class of one-step procedures for the numerical integration of nonlinear stiff
ordinary differential equations and differential-algebraic equations are implicit Runge-Kutta methods which
are characterized by pronounced stability and accuracy properties. The focus of this contribution is on the
application of diagonally implicit Runge-Kutta (DIRK) methods. In the case of quasi-static problems, it was
shown in [4, 9] that DIRK-methods are highly accurate and stable. In this presentation, DIRK-methods
are applied to structural dynamics problems within the context of the method of vertical lines. As already
mentioned in first investigations of linear structural dynamics, see [8], some stiffly accurate DIRK-methods
have better accuracy and stability properties as the HHT-α method, see [6]. However, these methods do not
achieve the same accuracy as in the case of quasi-static problems. In dynamics, order reduction phenomena can
be observed when applying DIRK-methods, which will be demonstrated by simple one-dimensional examples.
Introduction
In this contribution, we show the applicability of diagonally implicit Runge-Kutta schemes (DIRK) for
the time discretization of structural dynamic problems.
1 Space and time discretization

The balance of linear momentum and constitutive models of evolutionary-type lead to an initial boundary-
value problem. The problem is solved by consistent application of the method of vertical lines. In the first
step, the space discretization using standard finite elements is carried out yielding either an ODE-system
of second order for the discretized equation of motion or a system of differential-algebraic equations
(DAE-system) if constraints are also considered. In the case of inelastic materials, a further part in form
of an ODE-system of first order results from the evolution equations of the internal variables which have
to be evaluated at all spatial integration points. In the subsequent step, the time discretization is carried
out. Here, higher-order accurate DIRK-methods are the method of choice. For the application of the
DIRK-approach on the semi-discretized ODE-system of second order, it must first be transformed into an
ODE-system of first order by exploiting the trivial relationship u̇(t) = v(t) ∈ Rnuu . The semi-discretized
ODE-system of first order F(t, y(t), ẏ(t)) ∈ Rm of the dimension m = 2nuu + nQ is defined by
     
M v̇(t) + gu (u, q, t) + M ü(t) v(t) v0 
F(t, y(t), ẏ(t)) = u̇(t) − v(t) = 0, y(t) = u(t) ∈ Rm , y(t0 ) = u0 , (1)
     
q̇(t) − rQ (u, q, t) q(t) q0
with the unknown nodal displacements u(t) ∈ Rnuu , the unknown nodal velocities v(t) ∈ Rnuu , the known
nodal displacements u(t) ∈ Rnup , which are prescribed by appropriate Dirichlet boundary conditions,
and the internal variables q(t) ∈ RnQ , which describe the history dependence of the material behavior.
Applying the DIRK-approach, see [2, 3], to the ODE-system (1) yields in each stage a system of non-linear
equations
 S 
M Uni −
 Ûni
+ g (U , Q , T ) + M ü(T 
ni 
)
Gu (Uni , Qni ) ∆t 2
ā u ni ni ni
= n ii =0 ∈ Rnuu +nQ . (2)
L(Uni , Qni ) 
 Qni − Qni S 

∆tn aii − rQ (U ni , Q ni , Tni )
∗ Corresponding author: matthias.grafenhorst@tu-clausthal.de
133
The stage i, i = 1, . . . , s, is evaluated at the stage time Tni = tn + ci ∆tn . The system of non-linear
equations (2) depends on Pthe stage values Uni ∈ Rnuu , Qni ∈ RnQ , the time step-size ∆tn , the algorithmic
s
parameters aij , āij = k=1 aik akj and ci of the implicit s-stage Runge-Kutta method. The starting
S
Pi−1 S Pi−1
vectors Ûni = un + ci ∆tn vn + ∆t2n j=1 āij Anj and Qni = qn + ∆tn j=1 aij Q̇nj depend on the known
quantities un ∈ Rnuu , vn ∈ Rnuu and qn ∈ RnQ at time tn and the known stage derivatives Ani = (Uni −
S S
Ûni )/(∆t2n āii ), Q̇ni = (Qni − Qni )/(∆tn aii ) of the previous stages. After solving (2) for each stage i,
i = 1, . . . , s, the approximations at time tn+1 are computed by
s
X s
X s
X
un+1 = un + ∆tn vn + ∆t2n b̄i Ani , vn+1 = vn + ∆tn bi Ani , qn+1 = qn + ∆tn bi Q̇ni , (3)
i=1 i=1 i=1
Ps
with weighting factors bi , b̄i = j=1 bj aji . Owing to the fact that stiffly accurate, diagonally implicit
Runge–Kutta methods of higher-order are used, where the conditions asj = bj as well āsj = b̄j are
implied, the last calculation step (3) can be omitted. In particular, the stiff accuracy property ensures
un+1 = Uns , vn+1 = Vns , qn+1 = Qns .
2 Numerical results
In this section, we analyze the continuum oscillations in a one-dimensional elastic rod, see Fig. 1a, and
in a one-dimensional non-linear viscoelastic rod, see Fig. 1b. For both geometrically linear problems, the
u(x, 0) = 0 mm, u(x, 0) = 0 mm, u̇(x, 0) = 0 mm s−1 ,

u̇(x, 0) = 0 mm s−1 , t ∈ [0 s, 10 s] εv (x, 0) = 0, t ∈ [0 s, 10−2 s]
−2
E, A, ̺ F (t) = F̂ sin(2πf t) E, Ê, η0 , s0 , A, ̺ F (t) = F̂ sin(2πf t)

L L
x x
(a) Linear elastic rod (b) Non-linear viscoelastic rod
Figure 1. One-dimensional simulation model: fixed boundary condition (u(0, t) = 0 mm) on the left and an applied
external force F (t) on the right
load is applied at the free end as a sinusoidal force F (t) = F̂ sin(2πf t), with the amplitude F̂ = 105 N
and the excitation frequency f = 102 Hz. The left end of the rod is clamped. The treatment of the
first initial boundary value problem, see Fig. 1a, which is subjected to the specific initial conditions
u(x, 0) = 0 mm and u̇(x, 0) = 0 mm s−1 and the boundary conditions u(0, t) = 0 mm and ∂u(L, t)/∂x =
ε̂ sin(Ωt), ε̂ = F̂ /(EA), requires the
p solution of the wave equation ∂ u(x, t)/∂t = c ∂ u(x, t)/∂x , with
2 2 2 2 2
the longitudinal wave speed c = E/%. For a specific computation, we have chosen a parameter set
of E = 2.1 × 105 N mm−2 , A = 2 × 103 mm2 , L = 103 mm, % = 8 × 10−9 t mm−3 , where E denotes the
Young’s modulus, A the cross-sectional area, L the rod length and % the mass density. For the illustrated
linear elastodynamic problem, the integration methods detailed in Table 1 are analyzed in view of their
Table 1. Applied stiffly accurate DIRK-methods

abbreviation method order p stages s
BE Backward-Euler 1 1
Alex2/Ell Alexander/Ellsiepen 2 2
CAA trapezoidal rule or CAA 2 2
Alex3/Cash Alexander/Cash 3 3
expected and achieved temporal convergence order for the longitudinal displacement field u(x, t) and the
longitudinal velocity field v(x, t) = u̇(x, t), see Fig. 2. The basis for the calculation of the global error
measures erru and errv is formed by the analytical reference solution
∞ !
2ε̂ X (−1)n ωn Ω Ω2
u(x, t) = sin (ωn t) − 2 sin (Ωt) sin (kn x) + xε̂ sin (Ωt), (4)
L n=1 kn2 ωn2 − Ω2 ωn − Ω2
with kn = π/(2L)(2n − 1), ωn = c kn and Ω = 2πf . The approach adopted here for setting up these
reference solution is based on the strategies outlined in [7] and uses the concept of separation of vari-
ables for solving the hyperbolic partial differential equation. The temporal convergence behavior of the
134
DIRK-methods is compared with another time integration class, namely the generalized-α methods (ab-
breviation: Gen-α, order: p = 2) with the algorithmic parameterization defined in terms of the desired
amount of high-frequency dissipation and expressed by means of the spectral radius ρ∞ = 0.8, and is
depicted in Fig. 2. For more details to the Generalized-α method, we refer to [3] and the literature cited
therein. The space discretization of the rod using standard finite elements is carried out with ne = 104
three-node line Q2-elements (quadratic line elements). For modeling the inertia properties of the rod,
we use a consistent mass matrix. In Fig. 2a, we report comparable to problems in plasticity, see [1], an
1e-01
1e-02 1e-01
erru
errv
1e-03
BE 1e-02 BE
Alex2/Ell Alex2/Ell
1e-04 Trapez/CAA Trapez/CAA
Alex3/Cash Alex3/Cash
Gen-α Gen-α
1e-05 1e-03
2e-06 5e-06 1e-05 2e-05 5e-05 2e-06 5e-06 1e-05 2e-05 5e-05
∆tn in s ∆tn in s
(a) Global error of the displacement field (b) Global error of the velocity field
Figure 2. Convergence analysis for the linear elastic rod - Global error erru and errv vs. step-size ∆tn
order reduction where the third order of the Alex3/Cash method is not attained. As seen in Fig. 2b,
where the global error for the velocity field is depicted, further unsatisfactory tendencies are observed.
Here, the higher-order convergence is not obtained and is limited to one. As a possible explanation, we
assume that the hyperbolic solution characteristics inherent to the wave equation leads to a significant
degeneration of the temporal convergence behavior.
The modeling of the non-linear viscoelastic rod as shown in Fig. 1b utilizes the three-parameter model
of small strain viscoelasticity incorporating non-linear rate-dependence, see [5]. This step transforms
the initial boundary-value problem to be solved into the form ∂N (x, t)/∂x = %A∂ 2 u(x, t)/∂t2 , N = σA,
σ = Eε + Ê(ε − εv ), ε = ∂u/∂x, ε̇v = Ê/η(ε − εv ), η = η0 e−s0 |σ−Eε| where the same initial- and
boundary conditions for the longitudinal displacement and velocity field are taken as in the linear elastic
case. The internal variables in this constitutive model is the viscous strain field εv (x, t), whose initial
condition is given by εv (x, 0) = 0. For a complete problem description, the material parameter set will be
extended by the material parameters Ê = E = 2.1 × 105 N mm−2 , η0 = 13 N s mm−2 and s0 = 1.1 MPa−1 ,
associated with the overstress part of the constitutive model. Due the non-linear rate-dependence of the
applied constitutive model induced by the non-linear viscosity function η(ε, εv ), there exists no analytical
solution in contrast to the previously discussed elastodynamic problem. In order to compute the global
error measures erru , errv and errq , we conduct a high-precision reference solution employing Alex3/Cash
method with a constant step-size of ∆tn = 10−6 s. The results of the temporal convergence study are
shown in Fig. 3. The presence of a physical damping mechanism inherently anchored in the material
model leads obviously – in comparison to the order investigations in Fig. 2 – to an improved temporal
convergence behavior. A more descriptive interpretation of the statements made in this section are
provided by Fig. 4, in which the structural responses of the elastic and viscoelastic calculation models
are compared. From Fig. 4a, it is clear that the oscillation behavior of the linear-elastic rod is more
pronounced. On the other hand, for the viscoelastic rod, as a result of the viscous damping properties of
the material, almost all high-frequency components are filtered out, see Fig. 4b.
In the case of quasi-static problems, it was shown in [4, 9] that classical stiffly accurate DIRK-schemes
are highly accurate and stable. These numerical observations will be reviewed for the one-dimensional
physically non-linear example discussed here. For this purpose, only the inertia properties of the non-
linear viscoelastic rod are switched off so that an elliptic boundary-value problem must be solved within
the context of the method of vertical lines. The results of the temporal convergence study for the quasi-
static case are shown in Fig. 5 and are in perfect agreement with the theory.
135
1e-01 1e-01
1e-02 1e-01
1e-02
erru
errv
errq
1e-03 1e-03
BE 1e-02 BE BE
1e-04 Alex2/Ell Alex2/Ell Alex2/Ell
Trapez/CAA Trapez/CAA 1e-04 Trapez/CAA
Alex3/Cash Alex3/Cash Alex3/Cash
1e-05 Gen-α Gen-α Gen-α
1e-03 1e-05
1e-05 1e-04 1e-05 1e-04 1e-05 1e-04
∆tn in s ∆tn in s ∆tn in s
(a) Global error of the displacement (b) Global error of the velocity field (c) Global error of the internal vari-
field ables
Figure 3. Convergence analysis for the non-linear viscoelastic rod - Global error erru , errv and errq vs. step-size
∆tn
u(x, t) in mm u(x, t) in mm
(a) Space-Time-Displacement diagram for the linear (b) Space-Time-Displacement diagram for the non-linear
elastic rod: analytical solution, see Eq. (4) viscoelastic rod: numerical reference solution
Figure 4. Comparison of the oscillation behavior between linear elastic and non-linear viscoelastic system
1e-02
1e-02
1e-03
1e-03
erru
errq
1e-04
1e-04
1e-05 BE BE
Alex2/Ell 1e-05 Alex2/Ell
Trapez/CAA Trapez/CAA
1e-06 Alex3/Cash Alex3/Cash
1e-06
1e-05 1e-04 1e-05 1e-04
∆tn in s ∆tn in s
(a) Global error of the displacement (b) Global error of the internal vari-
field ables
Figure 5. Convergence analysis for the non-linear viscoelastic rod for the quasi-static problem - Global error erru
and errq vs. step-size ∆tn
3 Conclusions
In this contribution, the accuracy of higher-order DIRK-methods was analyzed with the help of a one-
dimensional wave propagation example. It is shown that the higher-order accuracy of DIRK-methods,
as proposed in [4, 9] for the quasi-static case, is not one-to-one transferable into the regime of structural
dynamics. Undesirable side effects in the area of structural dynamics occur for DIRK-integrators in
the form of order reduction phenomena, which reduce the performance of the analyzed time integration
methods.
136
References
[1] P. Ellsiepen and S. Hartmann. Remarks on the Interpretation of Current Non-linear Finite-Element-
Analyses as Differential-Algebraic Equations. International Journal for Numerical Methods in Engi-
neering 51 (2001), pp. 679–707.
[2] Matthias Grafenhorst. “Zeitadaptive Finite-Elemente-Berechnungen thermomechanisch gekoppelter
Problemstellungen sowie Mortarkontakt”. PhD thesis. Clausthal-Zellerfeld: Institute of Applied Me-
chanics, Clausthal University of Technology, 2018.
[3] Matthias Grafenhorst, Joachim Rang, and Stefan Hartmann. Time-adaptive finite element simula-
tions of dynamical problems for temperature-dependent materials. Journal of Mechanics of Materials
and Structures 12.1 (2017), pp. 57–91.
[4] S. Hartmann. Computation in finite strain viscoelasticity: finite elements based on the interpretation
as differential-algebraic equations. Computer Methods in Applied Mechanics and Engineering 191.13-
14 (2002), pp. 1439–1470.
[5] S. Hartmann. A remark on the application of the Newton-Raphson method in non-linear finite
element analysis. Computational Mechanics 36.2 (2005), pp. 100–116.
[6] Hans M. Hilber, Thomas J. R. Hughes, and Robert L. Taylor. Improved numerical dissipation for time
integration algorithms in structural dynamics. Earthquake Engineering and Structural Dynamics 5.3
(1977), pp. 283–292.
[7] Kurt Meyberg and Peter Vachenauer. Höhere Mathematik 2: Differentialgleichungen, Funktionenthe-
orie, Fourier-Analysis, Variationsrechnung. 4th ed. Springer-Lehrbuch. Berlin, Heidelberg: Springer
Berlin Heidelberg, 2001.
[8] B. Owren and H. H. Simonsen. Alternative integration methods for problems in structural dynamics.
Computer Methods in Applied Mechanics and Engineering 122.1-2 (1995), pp. 1–10.
[9] Steffen Rothe, Ahmad-Wahadj Hamkar, Karsten J. Quint, and Stefan Hartmann. Comparison of
diagonal-implicit, linear-implicit and half-explicit Runge–Kutta methods in non-linear finite element
analyses. Archive of Applied Mechanics 82.8 (2012), pp. 1057–1074.
137
138

A higher-order energy-momentum scheme for a non-

isothermal two-phase dissipation model of fibrous com-
posites
1
Michael Groß ∗
, Julian Dietzsch1 and Chris Röbiger1
1 TU Chemnitz, Professorship of applied mechanics and dynamics, Reichenhainer Straße 70, D-09126 Chemnitz, Germany.
Abstract. Fiber-reinforced plastics are made of an isotropic polymer matrix reinforced with inorganic or
organic fibers. These composites become more and more important in light-weight structures. The major
contribution to the internal dissipation in these fibrous composites is due to the isotropic matrix material.
However, internal dissipation in the fibers has to be taken into account if Carbon or Kevlar fibers, respec-
tively, are applied, or by using organic fibers, which are applied in the automotive industry, for example.
The presented two-phase dissipation model is based on the multiplicative split of the deformation gradient
of the composite in an elastic and viscous deformation gradient, as well as on the multiplicative split of
the fiber deformation gradient in an elastic and viscous fiber deformation gradient. The corresponding vis-
cous evolution equations are derived by a principle of virtual power together with further mechanical and
thermal time evolution equations of a dynamic continuous problem. For instance, in the internal power,
we consider free energy functions depending on matrix and fiber invariants. In the external power, the
non-negative internal dissipation with respect to a positive-definite matrix viscosity tensor and a positive
fiber viscosity parameter, respectively, are introduced. The internal power also includes mixed fields for
the thermo-elastic matrix and fiber behaviour. In this way, we obtain a mixed stress-strain formulation,
which also avoids locking behavior due to the stiff fibers embedded in the incompressible matrix mate-
rial. Algorithmic terms in the virtual external power provide the energy-momentum scheme of this two-
phase dissipation model. We demonstrate the independent dissipative material behaviour due to the matrix
and the fibers in dynamic numerical examples with different mechanical and thermal boundary conditions.
Introduction
The derivation of a viscoelastic material model for fibers from a three-dimensional constitutive model
for isotropic materials is successful (see Reference [3]). Therefore, we derive from the constitutive model
of the matrix material in Reference [1] a new fiber viscoelasticity formulation. In Reference [1], the
formulation is based on the well-known multiplicative split of the deformation gradient in a viscous and
an elastic product tensor. Considering also deformations of fibers, we thus assume the product structure
F F = F eF F vF of the fiber deformation gradient F F := at ⊗ a0 with the elastic fiber deformation gradient
F eF := at ⊗ ā and the viscous fiber deformation gradient F vF := ā ⊗ a0 . These second-order tensors are
defined by the stretched fiber direction vector at in the tangent space Tx Bt at the current configuration
Bt , the fiber direction vector ā in the linear space VF (‘intermediate fiber configuration’) and the unit
direction vector a0 with ka0 k = 1 in the tangent space TX B0 at the initial configuration B0 (see Fig. 1).
The symbol ⊗ denotes a dyadic tensor product. In principle, we assume a generalized Maxwell element
with an elastic and a viscoelastic branch, but considering large deformations, we define the fiber free
energy ΨF = ΨFela + ΨFvis . The elastic free energy ΨFela := Ψ̂Fela (C̃F ) depends on the squared fiber stretch
C̃F := at · at = a0 F T · F a0 = C : a0 ⊗ a0 = C : A0 = C F : I (1)
with the structural tensor A0 := a0 ⊗ a0 , the right Cauchy-Green tensor C := F T F , the unity metric
tensor I in B0 , and the right Cauchy-Green tensor C F := F TF F F = [C : A0 ] A0 of the fibers. Note that
the superscript letter ‘T ’ denotes the transposition of a second-order tensor. The elastic fiber stretch
T −T −1 −T −1
CFe := C eF : Ī = [F eF ] F eF : Ī = [F vF ] F T F [F vF ] : Ī = [F vF ] C [F vF ] : Ī (2)
∗ Corresponding author: michael.gross@mb.tu-chemnitz.de
139
a0
X
TX B0 Tx Bt
FF
at
F vF x
F eF
B0 ā
fiber VF
Bt
Figure 1. Fiber configurations with tangent spaces.
determines the viscoelastic free energy, where Ī designates the unity metric tensor of the linear space VF .
−1 −T −1
Tensor algebraic manipulations then leads to the relation CFe = C T : [F vF ] [F vF ] = C [C vF ] : I,
v
where the second-order tensor C F denotes the viscous right Cauchy-Green tensor
T
C vF = [F vF ] F vF = [a0 ⊗ ā] · [ā ⊗ a0 ] = [ā · ā] A0 =: CFv A0 (3)
The scalar-valued field CFv denotes the new viscous internal variable of the fibers, and the inverse of the
−1 −1 −T
viscous right Cauchy-Green tensor of the fiber family is defined by [C vF ] := [F vF ] [F vF ] , where the
v −1 v −1
inverse [F F ] satisfies the relation a0 = [F F ] ā =: λ̄F [a0 ⊗ ā] · ā = λ̄F [ā · ā] a0 = λ̄F CFv a0 . Thus,
the multiplier λ̄F denotes the inverse of the viscous internal variable, and the inverse of the viscous right
−1
Cauchy-Green tensor takes the form (CFv )2 [C vF ] = [a0 ⊗ ā] · [ā ⊗ a0 ] = [a0 ⊗ a0 ] CFv = CFv A0 . Hence,
1 1 C̃F
CFe = C A0 : I = [C : A0 ] v = v (4)
CFv CF CF
is the relation for the elastic fiber stretch. In this way, we obtain the viscous free energy function
ΨFvis := Ψ̂Fvis (CFe ) = Ψ̂Fela (C̃F [CFv ]−1 ), which means we apply the same free energy function as in the
elastic branch, but with a different argument and generally different material constants. According to the
Clausius-Plank inequality and the definition of the entropy, the viscous internal dissipation in the fibers
takes the form
1 C̃˙ F
DFint := SF C̃˙ F − Ψ̇Fela − Ψ̇Fvis = C̃F SF − Ψ̇Fela − Ψ̇Fvis = MF LF − Ψ̇Fela − Ψ̇Fvis > 0 (5)
2 2 C̃F
where we introduced the Mandel-like fiber stress MF := C̃F SF , which is energy-conjugated to the fiber
strain rate LF . The time rate of the free energy pertaining to the elastic branch then takes the form
∂ΨFela C̃˙ F ∂ΨFela

Ψ̇Fela = 2 = 2 C̃F LF =: MFela LF (6)
∂ C̃F 2 ∂ C̃F
and the time rate of the free energy associated with the viscoelastic branch reads
" #
vis ∂ΨFvis ˙e e ∂ΨF
vis
C̃˙ F ĊFv
Ψ̇F = C = 2 CF − =: MFvis [LF − LvF ] (7)
∂CFe F ∂CFe 2 CF 2 CFv

Thus, the viscous dissipation in the fibers are given by DFint := MF − MFela − MFvis LF + MFvis LvF > 0.
The Clausius-Plank inequality is then fulfilled by setting MF := MFela + MFvis , and by defining the viscous
evolution equation
.
MFvis = ΣFv (8)
with ΣFv := VF LvF as viscous driving force associated with the viscosity constant VF > 0.
140
1 Higher-order energy-momentum time integration
The energy-momentum time integration of the viscous fiber evolution equation requires that the relation
Z tn+1 Z 1 ◦
ΨFvis
t
− ΨFvis
tn
= Ψ̇Fvis dt ≡ ΨFvis dα (9)
n+1
tn 0
is exactly satisfied on each time step [tn , tn+1 ], or on the normalized time interval [0, 1], respectively (cp.
Reference [1]). Here, we indicate by a superimposed ◦ the derivative with respect to α ∈ [0, 1]. We
satisfy this constraint by means of a new algorithmic Mandel stress M̄Fv . Note that in Reference [1],
the energy-momentum time integration of the viscous free energy of the matrix material is only based
on the algorithmic elastic Mandel stress, which modifies the equations of motion only. This leads in
the isothermal case to an unmodified viscous evolution equation of the matrix material. Motivated by
Reference [2], in which the viscous evolution equation is modified by a discrete derivative, we here modify
also the viscous Mandel stress of the fibers. In this way, the equations of motion as well as the viscous
evolution equation are modified. Note also that, in contrast to Reference [2], the modification in this
paper is higher-order accurate. We determine the algorithmic Mandel stress M̄Fv by using a separate
constrained variational problem with the Lagrange functional
Z 1
FFv (M̄Fv , µvF ) := µvF GFv (M̄Fv ) + FFv (M̄Fv ) dα (10)
0
with the Lagrange multiplier µvF as well as a function and functional, respectively, defined by
Z 1
(M̄Fv )2
FFv (M̄Fv ) := and GFv (M̄Fv ) := ΨFvis − Ψ vis
F tn − GvF (M̄Fv ) dα (11)
2 t n+1
0
with respect to the constraint function

GvF (M̄Fv ) := MFvis + M̄Fv [LF − LvF ] (12)
The minimization of the Lagrange functional F leads to the Euler-Lagrange equations
Z 1 Z 1
δFFv v v v v v δFFv
≡ M̄ F δ M̄ F dα − µF [LF − LF ] δ M̄ F dα = 0 ≡ GFv = 0 (13)
δ M̄Fv 0 0 δµvF
which, in turn, lead to the algorithmic Mandel stress M̄Fv := µvF [LF − LvF ] with the Lagrange multiplier
Z 1
ΨFvis
tn+1
− Ψ vis
Ftn − MFvis [LF − LvF ] dα
v
µF := Z 1 0
(14)
v 2
[LF − LF ] dα
0
According to Reference [1], we apply a mixed finite element formulation which leads to an independent
fiber stress SF . The corresponding fiber stress equation then includes the algorithmic viscous Mandel
stress, so that we arrive at the space-time weak form
Z tn Z
˙ ∂ΨF M̄Fv
δ C̃F 2 + S̄F + − SF dV dt = 0 (15)
tn+1 B0 ∂ C̃F C̃F
where the algorithmic fiber stress S̄F fulfills the gradient theorem (cp. Eq. (9)) with respect to ΨFela . This
fiber stress then determines the stress power in the equation of motion. Further, we obtain the weak form
of the new viscous evolution equation of the fibers, given by
Z tn Z
M̄Fv VF
δ ĊFv −YF − + Lv
dV dt = 0 (16)
tn+1 B0 2 CFv 2 CFv F
Note that a further advantage of the existence of the additional algorithmic stress M̄Fv besides the algo-
rithmic stress S̄F is, that each free energy part of the fiber family possesses its own algorithmic modifi-
cation. Hence, neglecting one part of the fiber free energy allows for the neglection of the corresponding
algorithmic stress term. This simultaneous modification of the equation of motion and the viscous evo-
lution equation of the fibers by means of the same algorithmic stress term leads to the same stable time
integration as in Reference [2], but now also in the case of higher order accuracy in time.
141
2 Numerical example
We consider a displacement controlled tensile test of a tension bar. The fibers lie along the bar axis. On
the top (blue patches in Fig. 2, left) and at the bottom, we fix the temperature and horizontal positions,
but move in vertical direction. We apply symmetry boundary conditions in the x-z- and y-z-plain (green
patches in Fig. 2, left). The initial conditions are homogenous. In Fig. 2, right, we show the fiber stress.
Figure 2. Left: Boundary conditions of the tension bar. Right: Fiber stress at tn = 0.75 s (maximum load time).
Acknowledgments
This research is provided by the ’Deutsche Forschungsgemeinschaft’ (DFG) under the grants GR 3297/4-1
and 3297/6-1. This support is gratefully acknowledged.
References
[1] Michael Groß and Julian Dietzsch. Variational-based locking-free energy–momentum schemes of
higher-order for thermo-viscoelastic fiber-reinforced continua. Computer Methods in Applied Me-
chanics and Engineering 343 (2019), pp. 631–671.
[2] Michael Groß, Matthias Bartelt, and Peter Betsch. Structure-preserving time integration of non-
isothermal finite viscoelastic continua related to variational formulations of continuum dynamics.
Computational Mechanics 62.2 (2018), pp. 123–150.
[3] Boumediene Nedjar. An anisotropic viscoelastic fibre–matrix model at finite strains: continuum for-
mulation and computational aspects. Computer Methods in Applied Mechanics and Engineering
196.9-12 (2007), pp. 1745–1756.
142

Implicit Runge-Kutta Schemes and Galerkin Methods

Similarities and Differences
1
Bettina Schröder ∗
and Detlef Kuhl1
1 Institute of Mechanics and Dynamics, University of Kassel, Mönchebergstraße 7, 34109 Kassel
Abstract. In the present paper the numerical realization of linear elastodynamics is depicted, focussing on the
usage of higher order accurate time discretization schemes. Thereby, an implicit Runge-Kutta Radau IIA
scheme and a continuous Galerkin method are compared, depicting their similarities and differences.
Introduction
Social as well as economic conditions, together with an increasing environmental conscience lead to as-
pirations for high performance products manufactured by resource efficient production strategies. One
example of the latter is the creation of functionally graded structures out of mono-materials by using an
integrated thermomechanical forming process, cf. [9]. For a proper predictability of the structures’ prop-
erties using simulation techniques, appropriate material models as well as adequate numerical schemes
have to be exploited. In this context the application of higher order accurate time integration schemes
plays an important role. In the following an implicit Runge-Kutta Radau IIA scheme and a continuous
Galerkin method are analyzed, whereby linear elasticity is taken as a model problem for simplicity.
1 Fundamental Equation for Linear Elastodynamics
In order to derive the fundamental equation for elastodynamics, the principle of Jourdain is used, cf. [6].
Therefore, in a first step, the balance of energy of a continuum mechanical body B with domain Ω and
surface Γ
P (u̇) = K̇ + Ė + P ∗ = 0, (1)
is formulated. It consists of the rate of the kinetic energy K̇, the rate of the internal energy Ė and the
power due to external forces P ∗ defined as:
Z Z Z Z
K̇ = ρ0 ü · u̇ dV, Ė = ε̇ : Ce : ε dV, P ∗ = − ρ0 u̇ · f dV − u̇ · t∗ dA. (2)
Ω Ω Ω Γ
Therein, the expression ρ0 refers to the body’s density, the variable f to the applied mass forces and t∗
to the external Cauchy stress vector. The displacement field is indicated by u, while u̇ as well as ü are
the corresponding first and second time derivatives. The associated strain tensor and its rate are denoted
by ε and ε̇. The quantity Ce refers to the elastic constitutive tensor. Additionally, appropriate initial
conditions u0 = u(t0 ) as well as u̇0 = u̇(t0 ) are chosen.
In a second step, it is assumed that the total power P of a system reaches a stationary point, cf. [1, 8],
demanding that
0 ∈ ∂Pu̇ (u̇) (3)
holds, where ∂(•) represents the subdifferential with regard to the respective quantity, cf. [7, 10].
∗ Corresponding author: schroederbettina@uni-kassel.de, kuhl@uni-kassel.de
143
Evaluating Eq. (3) yields the weak form of elastodynamics

Z Z Z Z
ρ0 δ u̇ · ü dV + δ ε̇ : Ce : ε dV − ρ0 δ u̇ · f dV − δ u̇ · t∗ dA = 0. (4)
Ω Ω Ω Γt∗
2 Numerical Realization of Linear Elasticity
In order to solve Eq. (4) for a general continuum mechanical body, numerical schemes are adapted. First
of all Eq. (4) is discretized in space using the finite element method. Therefore, the body’s domain Ω is
divided in NE subdomains assuming that on each element Eq. (4) has to be fulfilled. Furthermore, the
geometry, the primary variable u, its time derivatives and the virtual displacement rate δ u̇ as well as the
corresponding gradients are approximated by nodal values and Lagrange shape functions. Inserting
these approximations in Eq. (4), evaluating the appearing integrals by a Gauss-Legendre quadrature,
assembling all element equations and applying the fundamental lemma of variational calculus, results in
the equation of motion
Mü + Ku = R∗ , with u0 = u(t0 ), u̇0 = u̇(t0 ) (5)
on structural level. Therein, M indicates the mass matrix, K the stiffness matrix and R∗ the vector of
external loads, cf. [11].
To determine the unknown displacement field, a relation between the accelerations ü and the displace-
ments u has to be established, before the emanating linear system of equations can be solved. This can
be done exploiting distinct time integration schemes. Therefore, Eq. (5) is rewritten in a system of first
order equations in time
∗
M̄ẇ + K̄w = R̄ , with w0 = w(t0 ) (6)
where
M 0 0 K v ∗ R∗
M̄ = , K̄ = , w= , R̄ = (7)
0 I −I 0 u 0
holds with the identity matrix I. Additionally, the time period of interest [0, T ] is divided into a number
of NT time steps [tn , tn+1 ] with step size ∆t.
2.1 Stiffly Accurate Fully Implicit Runge-Kutta Schemes
One method to discretize Eq. (6) in time is to use stiffly accurate fully implicit Runge-Kutta schemes. In
this context, Eq. (6) is evaluated only at discrete points in time tni , so called stages which are determined
by
tni = tn + ci ∆t, (8)
with the coefficients ci ∈ [0, 1] and i = 1, ..., s, where s embodies the maximum number of stages. The link
between the primary variable vector w and its time derivative is established by exploiting the fundamental
theorem of calculus as well as by approximating the appearing integrals using the particular quadrature
rules
X s X s
wn+1 = wn + ∆t bi ẇni , wni = wn + ∆t aij ẇnj , (9)
i=1 j=1
with weighting coefficients bi and aij , cf. [3]. The indices represent the points in time, where the respective
quantity is calculated, e.g. wni = w(tni ). To provide a better overview of the chosen quadrature
parameters, they are conventionally sorted into a Butcher tableau, see Fig. 1. In what follows, only
the two stage Radau IIA scheme of Fig. 1(b) will be considered. This method is, due to the property
a2i = bi , stiffly accurate so that the last stage is identical to the end of the time step tns = tn+1 and
Eq. (9)1 is superfluous. Moreover, its matrix A, see Fig. 1(c), can be inverted. Hence, Eq. (9)2 can be
reformulated and inserted in Eq. (6) leading to

1 −1 ∗ 1
M̄t At + K̄t wt = R̄t + M̄t A−1
t wt,n , (10)
∆t ∆t
144
(a) c1 a11 a12 ··· a1s (b) (c)

1 5 1
c2 a21 a22 ··· a2s −
3 12 12
.. .. .. .. .. 3 1 c A
. . . . . 1 −
4 4
cs as1 as2 ··· ass bT
3 1
b1 b2 ··· bs −
4 4
Figure 1. (a) Implicit Runge Kutta method, (b) Radau IIA scheme, (c) Short notation of the Butcher tableau
whereby wt = [wn1 , wn2 ] represents the aggregated stage values and wt,n includes two copies of wn . M̄t
∗
as well as K̄t are the augmentations of the mass and the stiffness matrix. The variable R̄t refers to the
extended external load vector. Analogously, At is a proper extension of the coefficient matrix A taking
into account the problem’s number of degrees of freedom or rather the size of wn . For the proposed
Radau IIA scheme Eq. (10) transfers into
 
3 1 ∗
 2∆t M̄n1 + K̄n1 2∆t
M̄n1  wn1 R̄n1 wn 2M̄n1
 −9 5  = ∗ + (11)
M̄n2 M̄n2 + K̄n2 wn2 R̄n2 ∆t −2M̄n2
2∆t 2∆t
and the unknown field variables can be determined by solving the linear system of equations.
2.2 Continuous Galerkin Methods
An alternative way to discretize Eq. (6) in time is to use continuous Galerkin methods. The idea
behind these schemes is, to apply the finite element method to the time domain, cf. [2]. Thus, Eq. (6) is
multiplied with a test function δt and integrated over the time domain to yield the weak form
Z 1
∗
δt M̄ẇ + K̄w − R̄ Jt dξt = 0. (12)
−1
The integral therein is linked to a standard time element based on natural coordinates ξt ∈ [−1, 1],
exploiting
∆t tn + tn+1
tn = ξt + (13)
2 2 .
The term Jt = ∆t/2 denotes the Jacobian in time. To solve the weak form Eq. (12) in time the
primary variable and its time derivative are approximated employing nodal values and Lagrange shape
functions Nti (ξt ) of polynomial degree pt with i ∈ [1, pt + 1]. The test function is approximated employing
Lagrange shape functions N̄ti (ξt ) of polynomial degree p̄t with i ∈ [1, p̄t + 1]
pX
t +1 pX
t +1 p̄X
t +1
w≈ wi Nti (ξt ), ẇ ≈ wi Ṅti (ξt ), δt ≈ δti N̄ti (ξt ). (14)

i=1 i=1 i=1
Additionally, the integral in Eq. (12) is determined using a quadrature rule with NGt quadrature points
and weighting coefficients, so that at the end a linear system has to be solved. In the following special
choices for the shape functions and the quadrature rule will be depicted in detail. The idea is to elect
pt = NGt = p̄t + 1 = 2, to consider a Radau quadrature rule with quadrature points ξ̄ t = [−1/3 1] as
well as with weighting coefficients [3/2 1/2], to designate the Radau quadrature points as supporting
points for the Lagrange shape functions N̄ti (ξt ) and to choose ξ t = [−1 ξ̄ t ] as supporting points for the
Lagrange shape functions Nti (ξt ). With the interpolatory property of the Lagrange shape functions,
cf. [4], the following results are obtained
N̄t1 (−1/3) = 1 N̄t2 (−1/3) = 0 N̄t1 (1) = 0 N̄t2 (1) = 1

Nt1 (−1/3) = 0 Nt2 (−1/3) = 1 Nt3 (−1/3) = 0 (15)
Nt1 (1) = 0 Nt2 (1) = 0 Nt3 (1) = 1.
145
The evaluations of the shape functions’ derivatives result in

2 6 2
Ṅt1 (−1/3) = − Ṅt2 (−1/3) = Ṅt3 (−1/3) =
∆t 4∆t 4∆t (16)
2 18 10
Ṅt1 (1) = Ṅt2 (1) = − Ṅt3 (1) =
∆t 4∆t 4∆t .
Inserting Eq. (14), Eq. (15) and Eq. (16) into Eq. (12) leads to
   
3 ∆t 1 " 2# ∆t ∗ " #
 4 M̄(−1/3) + K̄(−1/3) M̄(−1/3)  w  2 R̄ (−1/3)  1 −M̄(−1/3)
 2 4 = +w (17)
−9 5 ∆t  w3 ∆t ∗ −M̄(1)
M̄(1) M̄(1) + K̄(1) R̄ (1)
4 4 2 2
and the unknown field variables can be determined by solving the linear system of equations. Com-
paring Eq. (17) to Eq. (11) shows that both forms are identical if Eq. (17) is multiplied by 2/∆t and
if tn (−1/3) = tn1 as well as tn (1) = tn2 are taken into account. Hence, in the case of linear elasticity
the fully implicit Runge-Kutta scheme can be interpreted as a special continuous Galerkin method,
although the derivation strategy is a different one.
3 Conclusion and Outlook
In the present paper, the fundamental equation of linear elastodynamics is derived and its numerical
realization is demonstrated. The focus is laid on the time discretization, using a two stage implicit
Radau IIA scheme as well es a continuous Galerkin method with Radau quadrature rule and La-
grange shape functions. It could be shown, that in contrast to [5], both time integrators yield, in
the case of linear elastodynamics, the same system of linear equations and, hence, can be considered as
identical. Nevertheless, the methods’ derivation strategies are different, except that both use quadrature
rules for approximating integrals.
Open questions, however, are whether this behavior can be transferred to the nonlinear case and if the
identity between implicit Radau IIA and continuous Galerkin schemes can be shown for higher stage
orders as well. A further issue, which has to be clarified is - what are the consequences, if systems of
second order in time are treated directly without reduction to first order systems.
References
[1] P. Germain. The method of virtual power in continuum mechanics. Part 2: Microstructure. SIAM
Journal on Applied Mathematics 25 (1973), pp. 556–576.
[2] T. Gleim and D. Kuhl. Higher order accurate discontinuous and continuous p-Galerkin methods for
linear elastodynamics. Zeitschrift für angewandte Mathematik und Mechanik 93 (2013), pp. 177–
194.
[3] E. Hairer and G. Wanner. Solving Ordinary Differential Equations II. Springer-Verlag, 2002.
[4] T. J. R. Hughes. The Finite Element Method. Linear Static and Dynamic Finite Element Analysis.
Dover Publications, INC., 2000.
[5] B. L. Hulme. Discrete Galerkin and Related One-Step Methods for Ordinary Differential Equations.
Mathematics of Computation 26 (1972), pp. 881–891.
[6] P. E. B. Jourdain. Note on an Analogue of Gauß’ Principle of Least Constraints. Quarterly Journal
of Pure and Applied Mathematics 40 (1909), pp. 153–157.
[7] R. T. Rockafellar. Convex Analysis. Princeton University Press, 1970.
[8] B. Schröder and D. Kuhl. Small strain plasticity: classical versus multifield formulation. Archive of
Applied Mechanics 85 (2015), pp. 1127–1145.
[9] K. Steinhoff, U. Weidig, and N. Saba. Investigation of Plastic Forming Under the Influence of Locally
and Temporally Variable Temperature and Stress States. In: Functionally Graded Materials in
Industrial Mass Production. Ed. by K. Steinhoff, H.J. Maier, and D. Biermann. Auerbach: Verlag
Wissenschaftliche Scripten, 2009.
[10] E. Zeidler. Nonlinear Functional Analysis and its Applications III. Springer Verlag, 1985.
[11] O. C. Zienkiewicz and R. L. Taylor. The Finite Element Method. Butterworth Heinemann, 2000.
146

Higher order reciprocal mass matrices via heuristic cus-

tomization
Anton Tkachuk ∗ 1
1 Institute for Structural mechanics, University of Stuttgart, Pfaenwaldring 7, 70550 Stuttgart, Germany
Abstract. Customization of nite elements for low dispersion error through grid dispersion analysis requires
a symbolic expansion of a determinant of a representative dynamic stiness matrix. Such an expansion turns
out to be a bottleneck for many practical cases with the size of the representative matrix greater than eight
or ten even if the modern computer algebra systems are applied. In this contribution, we propose an alter-
native approach for low-dispersion customization that avoids explicit determinant expansion. This approach
reduces the customization problem to a series of quadratic programming problems and consists of two main
steps. First, the customization problem is reformulated as a rank minimization problem for the representative
dynamic stiness matrix evaluated at several discrete pairs of wavenumbers and frequencies. Second, the
rank minimization problem is solved approximately via log-det heuristics. An example for customization of
reciprocal mass matrices veries the proposed approach.
Introduction
Several formulations of reciprocal mass matrices are proposed recently in [2, 3, 6, 8] and they contain
naturally few free real-valued parameters. These free parameters are chosen to meet various performance
goals for transient structural problems, like low dispersion error at acoustic branches and higher critical
time step. The process of such parameter selection is called customization.The former goal is usually
achieved through Grid Dispersion Analysis (GDA), i.e. obtaining relation between angular frequency and
wavevector for a propagating wave in a given discretization. Necessary conditions for low dispersion error
are imposed on the coecients of Taylor expansion of the dispersion relation yielding a set of algebraic
equations with respect to the free parameters. These favorable values of the free parameters increase the
convergence order beyond the expected performance of the consistent mass matrix and it leads to so-
called high order reciprocal mass matrices, see examples [6, 8] . Unfortunately, symbolic evaluation of the
dispersion relation is not always possible because of a determinant of the representative dynamic stiness
matrix should be expanded symbolically in an intermediate step of GDA. Such a symbolic expansion is
not tractable for many practical cases with the size of the representative matrix larger than eight or ten
even if the modern computer algebra systems are applied. Therefore, an alternative approach is sought
that avoids explicit expansion of the determinant.
A conceptually new way for dispersion customization is based on enforcing not full rank of the represen-
tative dynamic stiness matrix instead of setting its determinant to zero [7]. Symbolic rank evaluation
is still a hard problem. Therefore, it is solved for the matrix evaluated at few pairs of angular frequency
and wavevector. The latter problem belongs to a class of well-studied ane rank minimization problems
(ARMP) [1, 4] and it is solved numerically by log-det heuristic [1]. Anity means that the underlying
representative dynamic stiness matrix linearly depends on the free parameters, which is the case for
mentioned above formulations of reciprocal mass matrix.
Consider a representative dynamic stiness matrix (RDSM)
Kdyn,rep (s, ω, k) = QC◦ (s)KS(k) − ω 2 Irep (1)
∗ Corresponding author: tkachuk@ibb.uni-stuttgart.de ORCID iD: 0000-0001-8572-8532
147
for a given representative patch with dimension nrep , wave vector k and angular frequency ω . Irep denotes
an identity matrix of size nrep . RDSM depends on the stiness matrix K, a reciprocal mass matrix C◦ ,
ltering matrix Q and a shift matrix S(k) as shown in [7]. The latter two matrices describe topology
and periodicity of the representative patch.
The proposed problem statement avoids explicit symbolic determinant expansion of RDSM. Instead it
assumes some knowledge about the branches of the dispersion relation of the continuous model that
should be accurately modeled. We select nw frequency-wavevector pairs (ωj , kj ) j = 1, nw that satisfy
the continuum dispersion relations and enforce the singularity of the matrix in Eq. (1) at these pairs in
form
rank (Kdyn,rep (s, ωj , kj )) < nrep , j = 1, nw . (2)
Checking all nw inequalities is still a hard problem. Instead a good solution is sought that minimizes the
sum of the ranks of RDSM for all pairs j. Such a sum now denes the objective of the customization.
The optimization problem has constraints on the plausible values for the free parameters. A convex
subset s∈K arises naturally in the construction of the reciprocal mass matrices, where all customization
parameters are non-negative and the sum of them is less than one due to positive deniteness restriction on
RMM suggested in [5, 8]. Having the objective function and constraints at hand, the problem statement
can be given.
Problem statement. Given a set of pairs frequency, wavevector (ωj , kj ) with j = 1, nw and a repre-
sentative dynamic stiness matrix with dimension nrep that linearly depend on a set of free parameters
s = [sj ] with j = 1, np . Find the values of the free parameters s∗ from a convex subset K such that
" #
∗
s = arg min rank diag (Kdyn,rep (s, ωj , kj )) . (3)
s∈K j=1,nw
Here, an operator diag denotes a constructor of a block diagonal matrix from a list of rectangular matrices.
The stated above ARMP is solved approximately by log-det heuristic [1]. A substantial dierence to the
original method is due complex entries of the RDSM matrices. The original rank minimization algorithm
is valid for a real semi-denite matrices. To bring our problem to an appropriate shape, a matrix X is
introduced with

X = diag KH
rep,dyn
(s, ωj , kj )Krep,dyn (s, ωj , kj ) (4)
j=1,nw
with superscript H being Hermitian transpose. The resultant matrix is a Hermitian positive semi-
denite matrix, which is sucient for the original log-det heuristics [1]. The rank of matrix X coincides
with the objective function in Eq. (3), but dependency on the free parameters is quadratic. Now, a rank
minimization problem is formulated for matrix X with
s∗ = arg min (rank(X)) (5)

s∈K
Rather than solving rank minimization problem, a log-det smooth surrogate for rank(X) is used
s∗ = arg min (log det (X + δI)), (6)

s∈K
where δ > 0 is a small regularization constant. Despite complex entries of the matrix X, the determinant
of Hermitian positive denite matrix (X+δI) is real and positive. Here, we follow the paper [1] and reduce
the optimization to a series of quadratic programming problems. A Taylor expansion of the objective
function in the vicinity of current value Xk is carried out
log det(X + δI) ≈ log det(Xk + δI) + trace (Wk (X − Xk )) , (7)
where Wk = (Xk + δI)−1 is a weighting matrix and it is computed numerically. The rst term in
the linearization does not inuence extremum and it is never computed. The second term is quadratic
with function with respect to customization parameters s and it is minimized using standard quadratic
programming algorithm on a convex set K. This brings us to the customization algorithm described in
Figure 1.
148
1. Given regularization parameter δ > 0, tolerance >0 and number of iteration maxiter
2. Initialize counter k=0 and free parameters s0 ∈ K,

3. Compute X0 = X(s0 )
4. Repeat until k< maxiter
5. Compute weighting Wk = (Xk + δI)−1

6. Solve QP: sk+1 = arg mins∈K trace (Wk (X − Xk ))
7. Compute Xk+1 = X(sk+1 )

8. If ||sk+1 − sk ||2 < then break;
9. Goto 4.
Figure 1. Algorithm for customization based on log-det heuristic proposed in [1].
2 Example
An example for the proposed customization approach is presented for a 6-node triangular acoustic element
in 2D. Only uniform hexagonal mesh with element size h with a uniform isotropic material (phase velocity
c0 = 1, ρ = 1.0 and Kbulk = 1.0) is considered. An implementation of the proposed algorithm in computer
algebra system Maple is used for all examples. Accuracy of the computations is controlled by assigning
a value to the environment variable Digits to 40. Quadratic problems within algorithm are solved by a
procedure Minimize of the package Optimization. We use the default options of the procedure.
s = [C21 , C22 ] within a convex set

This element has two free parameters for the reciprocal mass matrix
K := {C21 ≥ 0; C22 ≥ 0; 1 − C21 − C22 ≥ 0}. GDA predicts an optimal low-frequency accuracy in case
∗
of hexagonal meshes for values s = [0.36, 0.41]. These values provide fourth order of accuracy with
small constants at sixth order error term in the acoustic branch. This case studies an over-determined
set (nw > ns ) with of four frequency-wavevectors pairs
√ √
hk1 = [0.1, 0], hk2 = [0.2, 0], hk3 = [0.05 3, 0.05], hk4 = [0.1 3, 0.1], ωj = c0 |kj |, (8)
with wavevectors along x−axis and in the direction tilted by 30◦ to x−axis. The size of the representative
−13
patch nrep is four. Thus, the dimension of the matrix X is 16. We use a tolerance = 10 , regularization
−24
δ = 3 · 10 , 101 iterations and initialization of customization parameters s0 = [0.34, 0.45] in this case.
The algorithm converges within eleven iterations to a solution
∗ ∗
C21 = 0.3599102881, C22 = 0.4334932521 (9)
with the lowest eigenvalues of RDSM below 10−15 .
3 Conclusions
A novel method for customization of reciprocal mass matrices is presented that avoids symbolic expansion
of the determinant of the representative dynamic stiness matrix. The customization problem is recast
into a rank minimization problem for the representative dynamic stiness matrix at discrete frequency-
wavenumber pairs. The rank minimization problem is consequently solved via det-log heuristic [1]. The
log-det heuristic provides a smooth surrogate function of the matrix rank whose minimum can be nd by
nonlinear programing algorithms. Here, this minimum is found by sequence of quadratic programming
solves as suggested in [1]. An example veries applicability of the approach to a customization of a 6-node
triangular acoustic element in 2D.
149
Acknowledgments
The work of the author was supported by the Deutsche Forschungsgemeinschaft (DFG - German Research
Foundation) Projektnummer 326748051 (TK 63/1-1).
References
[1] Maryam Fazel, Haitham Hindi, and Stephen P Boyd. Log-det heuristic for matrix rank minimization
with applications to Hankel and Euclidean distance matrices. Proceedings of the 2003 American
Control Conference. 3 (2003), pp. 21562162.
[2] José A González, R Kolman, SS Cho, CA Felippa, and KC Park. Inverse Mass Matrix via the
Method of Localized Lagrange Multipliers. International Journal for Numerical Methods in Engi-
neering (2017), pp. 124.
[3] Jose A. Gonzalez, Jan Kopa£ka, Radek Kolman, S. S. Cho, and K. C. Park. Inverse Mass Ma-
trix for Isogeometric Explicit Transient Analysis via the Method of Localized Lagrange Multipliers.
International Journal for Numerical Methods in Engineering 117 (2019), pp. 939966.
[4] Karthik Mohan and Maryam Fazel. Iterative reweighted algorithms for matrix rank minimization.
Journal of Machine Learning Research 13.Nov (2012), pp. 34413473.
[5] Anne-Kathrin Schaeuble. Variationally Consistent Inertia Templates for Speed-up and Customiza-
tion in Explicit Dynamics. PhD thesis. Stuttgart: Institut für Baustatik und Baudynamik der Uni-
versität Stuttgart, 2019.
[6] Anne-Kathrin Schaeuble, Anton Tkachuk, and Manfred Bischo. Variationally consistent inertia tem-
plates for B-Spline and NURBS-based FEM: Inertia scaling and customization. Computer Methods
in Applied Mechanics and Engineering 326 (2017), pp. 596621.
[7] A Tkachuk. Customization of reciprocal mass matrices via log-det heuristic (submitted). Interna-
tional Journal for Numerical Methods in Engineering (2019).
[8] A. Tkachuk and M. Bischo. Direct and sparse construction of consistent inverse mass matrices:
general variational formulation and application to selective mass scaling. International Journal for
Numerical Methods in Engineering 101.6 (2015), pp. 435469.
150
MS 08: Smart, Active and Polymer materials

Modeling the thermomechanical behavior of strain-induced

crystallization in unlled polymers
Serhat Aygün ∗ and Sandra Klinge
Institute of Mechanics, TU Dortmund University, Leonhard-Euler-Strasse 5, 44227 Dortmund, Germany
Abstract. The crystallization in certain polymers, such as natural rubber, is a phenomenon manifesting it-
self as the natural reinforcement caused by high deformation. Starting from crystallization nuclei, polymer
chains leave their natural entangled structure, stretch out, fold back and stack to regions with a regular
structure. This property of polymers must be taken into consideration when planning manufacturing pro-
cesses, since it signicantly inuences mechanical and thermal properties of the nal product. In the present
contribution, the described material behavior is simulated by a coupled thermomechanical model. The mi-
cromechanical continuum model, involving the degree of network regularity as an internal variable, has the
following features: First, the thermodynamically consistent framework enables us to simulate the reduction of
the crystallinity degree during the unloading phase by proposing a dissipation potential and by assuming a
specic coupling condition between the inelastic deformations and the network regularity. Secondly, possible
material microstructures are reconstructed on the basis of experimental data, and the evolution of the net-
work regularity under the cyclic tensile load is visualized. In addition to the purely mechanical framework,
the heat conduction equation is solved in order to simulate the temperature evolution. The resulting, thermo-
mechanically coupled system of equations is solved in a monolithic way. Finally, selected examples for unlled
natural rubber compare the numerically simulated eective material response with the experimental data.
Introduction
The phenomenon of strain-induced crystallization (SIC) in polymers is characterized by the fact that,
under the inuence of high strains, polymer chains partially form a regular structure that resembles a crys-
talline pattern. Under the increasing strain, the evolution of the microstructure causes the strengthening
of material and changes the reaction to the macroscopic mechanical and thermal inuences. Experimental
data for cyclic uniaxial tests show that the stress response builds a hysteresis which indicates that the
process has a dissipative character. Moreover, the results show that crystallization starts at a stretch
of about 430%. At the stretch of 600% the crystallinity degree reaches 15%. The higher values of crys-
tallinity can be achieved if the higher stretches are applied. However, additional inelastic processes occur
simultaneously to the SIC in that case. It is also worth mentioning that heat produced by the formation
of crystalline regions leads to an increase in temperature, whereas the regression of crystalline regions
during the unloading phase yields a temperature decrease. Accordingly, the temperature diagram has an
analogous form to the diagram of crystallinity degree as shown in [2].
Most of the models focusing on the mechanical modeling of polymers use the Langevin expression for the
free energy as a basis [3, 6]. However, the model presented here treats a polymer aected by the strain
induced crystallization as a heterogeneous medium consisting of regions with a dierent degree of network
regularity. The model proposed is thermodynamically consistent. It is based on the Arruda-Boyce as-
sumption for the Helmholtz free energy and on the assumption for the dissipation potential [1]. The free
energy primarily includes terms due to elastic deformation and temperature. The external variables are
temperature and deformation, whereas the inelastic deformations and degree of the network regularity
are internal variables. Their evolution equations are derived according to the minimum principle of the
dissipation potential [4]. In addition to the balance of linear momentum describing the mechanical part,
the heat equation is used to simulate the temperature response due to the crystallization.
∗ Corresponding author: serhat.ayguen@tu-dortmund.de
151
1 Micromechanical modeling of SIC in polymers
The theory of nite deformations is the suitable framework for the simulation of polymers. Typical of this
theory, the modeling of dissipative processes requires a multiplicative decomposition of the deformation
gradient into an elastic part and an inelastic part F = F e ·F c . In our case, F c represents the deformations
caused by the crystallization. However, the multiplicative decomposition alone is not enough to describe
the phenomena such as the formation and growth of crystalline regions. For this purpose, an internal
variable is introduced denoting the regularity of the polymer chain network χ. This variable takes values
from range [0,1], where 0 corresponds to a totally amorphous polymer and 1 to completely crystalline
regions. The minimum principle of the dissipation potential is pursued in order to derive the evolution
laws obeying the second thermodynamic law. According to this theory, the Lagrange function involves
the elastic power and dissipation potential L = Ψ̇ + ∆. In this work, the Helmholtz energy Ψ(F e , Θ) =
Ψe (F e ) + ΨΘ (Θ) consists of an elastic part depending on elastic deformations and a purely temperature
dependent part, where Θ is the temperature. We assume the free energy of an Arruda-Boyce material
for the elastic part with a split into a volumetric and a deviatoric contribution

µ̄ ¯ m 2 11 m2 3
e e e e e
Ψ (F ) = Ψ (J , C ) = K U (J ) + e
I1 − 3 + ¯
I −9 + ¯
I1 − 27 ,
2 10 1 525
(1)
1 −2 µ 1
U (J e ) = (J e )2 − 1 − 2 ln(J e ) , I¯1 = J e 3 tr(C e ) , µ̄ = , m = .
4 99
1 + 35 m + 175 m2 λ2m
T
In the chosen constitutive law, C e = (F e ) · F e represents the elastic right Cauchy-Green deformation
tensor, J e = det (F e ) is the elastic Jacobian, K is the bulk modulus, µ denotes the shear modulus and
λm is the limiting network stretch. We furthermore assume the dissipation potential and the evolution
law for variable B

c ˙ tr(M )
∆ = (A + B) |χ̇| , Ḃ = sgn kM dev k |χ̇| , M dev = M − I , (2)
f (χ) 3
where the material constant A is the crystallization limit, being a threshold for the start of the regularity
evolution and B determines its evolution depending on the stress state. Symbol M dev denotes the
deviatoric part of the Mandel stress tensor M = F e T · ∂Ψ/∂F e . In the present model, the regularity
evolution is favored in regions with a high value of the regularity degree which is achieved by introducing
function f (χ). Finally, the evolution of the F c is coupled with the evolution of regularity χ through the
relationship for velocity gradient Lc

c c c −1 ˙dev M dev
L = Ḟ · F = k χ̇ sgn kM k , (3)
kM dev k
where k is a positive proportionality constant. The particularity of the assumption (3) is the introduction
of the signum function depending on the evolution of the Mandel stress tensor. This enables the regularity
to increase during the loading phase and to subsequently decrease during the unloading phase.
2 Thermodynamic consistency and thermal modeling of SIC
The study of thermal eects starts with the consideration of the dissipation inequality which is expressed
by using the Clausius-Duhem formalism
1 1
D = −Ψ̇ − η Θ̇ + P : Ḟ − q · ∇X Θ ≥ 0 . (4)
ρ0 ρ0 Θ 0
This formulation of the second thermodynamic law depends on the density in reference conguration
(ρ0 ), the rst Piola-Kirchho stress tensor (P ), the entropy (η ) and the heat ux vector (q 0 ). The time
derivative of the free energy now reads
∂Ψ e ∂Ψ
Ψ̇ = : Ḟ + Θ̇ , F˙ e = Ḟ · F c −1 − F e · F˙ c · F c −1 , (5)
∂F e ∂Θ
such that Eq. (4) is reformulated as

1 ∂Ψ c −T c ∂Ψ 1
D= P− · F : Ḟ + M : L − + η Θ̇ − q · ∇X Θ ≥ 0 . (6)
ρ0 ∂F e
∂Θ ρ0 Θ 0
152
Inequality (6) yields two consequences: Firstly, the constitutive laws for the the rst Piola-Kirchho
stress tensor P = ρ0 ∂Ψ/∂F e · F c −T and the entropy η = −∂Ψ/∂Θ are obtained by assuming that
each expression in parentheses is equal to zero. Secondly, the dissipation includes a term due to the
crystalization (DSIC ) and a term due to heat conduction (Dcond ). As usual, the positivity of each term
is required separately:
1
DSIC + Dcond ≥ 0 , DSIC = M : Lc = qχ χ̇ ≥ 0 , Dcond = − q · ∇X Θ ≥ 0 . (7)
ρ0 Θ 0
Dissipation DSIC implies that the regularity evolution and the corresponding thermodynamic driving
force (qχ ) have the same sign, whereas the classic Fourier's conduction law is introduced as a solution of
inequality (7 c). The thermal problem is described by the local form for the balance of energy
1 1 ∂2Ψ
Ψ̇ + η Θ̇ + Θ η̇ + Div(q 0 ) = P : Ḟ + rΘ , η̇ = − 2 Θ̇ , (8)
ρ0 ρ0 ∂Θ
where rΘ is the heat source. Finally, the insertion of Eq. (5) into Eq. (8) yields the heat conduction
equation
1 ∂2Ψ
cd Θ̇ + Div(q 0 ) = rΘ + qχ χ̇ , cd = −Θ 2 , (9)
ρ0 ∂Θ
determining the temperature change caused by the SIC. At this stage, our model assumes the heat
capacity (cd ) to be constant during the process.
3 Numerical example
The material model for SIC is implemented into the FE program FEAP [7]. In order to illustrate its
application, a uniaxial tensile test for a square sample with the dimensions 100 × 100 nm is chosen. The
lower and the upper boundaries are constrained and a gradually increasing displacement within the range
[0, 250 mm] is applied at each of these boundaries in the normal direction. Thereafter, the specimen is
gradually unloaded. The total loading time amounts to 10 s and the time increment is ∆t = 1e−3 s. The
regularity of the polymer network in the starting conguration has random initial values within the range
[0, 1e−3], which simulates the nuclei of the crystalline regions (Fig. 1 a). During the loading phase the
regularity increases and its growth is favored in regions with the higher initial values. This is achieved by
introducing function f (χ) in the evolution of B (Eq. (2 b)). Figure 1 b shows the state of the network
regularity in the nal loading step. At this stage, the crystallinity degree is about 18 %. Subsequently,
the crystalline regions decrease with a lower rate during the unloading phase (results not shown here).
Finally, the material becomes entirely amorphous.
2.22E-16 4.05E-08
5.94E-05 0.00E+00
1.19E-04 9.00E-02
1.78E-04 1.80E-01
2.38E-04 2.70E-01
2.97E-04 3.60E-01
3.56E-04 4.50E-01
4.16E-04 5.40E-01
4.75E-04 6.30E-01
5.35E-04 7.20E-01
5.94E-04 8.10E-01
6.54E-04 9.00E-01
7.13E-04 1.00E+00
a) b)
Time = 2.00E-03 Time = 5.00E+00
Figure 1. a) The initial microstructure with a random distribution of the regularity. b) Distribution of the
regularity in the nal loading step. The applied material parameters are: K = 5e8 Pa, µ = 4e5 Pa,
λm = 3, k = 7e−2, A = 1e6 Pa, c = 2e5 Pa and cd = 2.2e5 J/(kg K).
The temperature distributions during the loading phase at a stretch of λ = 5 and in the last loading
step at a stretch of λ = 6 are shown in Fig. 2. The temperature development behaves similarly to the
regularity evolution: The highest temperature occurs in the completely crystallized regions, whereas the
temperature in the amorphous regions hardly changes during the process.
153
2.71E+02 2.66E+02
2.70E+02 2.70E+02
2.72E+02 2.72E+02
2.74E+02 2.74E+02
2.76E+02 2.76E+02
2.78E+02 2.78E+02
2.80E+02 2.80E+02
2.82E+02 2.82E+02
2.84E+02 2.84E+02
2.86E+02 2.86E+02
2.88E+02 2.88E+02
2.90E+02 2.90E+02
2.78E+02 2.93E+02
a) b)
Time = 4.00E+00 Time = 5.00E+00
Figure 2. a) Temperature distribution (in Kelvin) during loading at a stretch of λ = 5. b) Temperature dis-
tribution (in Kelvin) in the nal loading step. The applied material parameters are the same as in
Fig. 1.
4 Conclusions
The present work shows a thermodynamically consistent model based on the assumptions for the free
energy and dissipation potential. The main advantages of the approach are: i) It enables the simulation
of the formation and development of crystalline regions during loading and degradation of crystalline
regions during unloading. ii) The microstructure evolution is visualized, which is not possible with
experimental techniques. iii) The heat production due to the SIC and the related temperature change
can be observed. iv) The model is applicable for implementation within the multiscale concept [5] in
order to study the inuence of crystalline regions on the eective material behavior of polymers. The
proposed model is suitable for further extensions and oers a good basis for simulating eects, such as
damage, crack propagation and coupling with plastic deformations. In all these cases, a comparison with
additional experimental results would be necessary for a reliable determination of material parameters.
Acknowledgments
The authors, S. Aygün and S. Klinge, gratefully acknowledge the nancial support of the German Research
Foundation (DFG), research grant No. 324689375.
References
[1] S. Aygün and S. Klinge. Study of the microstructure evolution caused by the strain-induced crystal-
lization in polymers. PAMM 18 (2018), e201800224.
[2] N. Candau, R. Laghmach, L. Chazeau, J.-M. Chenal, C. Gauthier, T. Biben, and E. Munch. Inuence
of strain rate and temperature on the onset of strain induced crystallization in natural rubber. Eur.
Polym. J. 64 (2015), pp. 244252.
[3] R. Dargazany, V. N. Khiêm, E. A. Poshtan, and M. Itskov. Constitutive modeling of strain-induced

crystallization in lled rubbers. Phys. Rev. E 89 (2014), p. 022604.
[4] K. Hackl and F. D. Fischer. On the relation between the principle of maximum dissipation and
inelastic evolution given by dissipation potentials. Proc. R. Soc. A 464 (2008), pp. 117132.
[5] S. Klinge and K. Hackl. Application of the multiscale FEM to the modeling of nonlinear composites
with a random microstructure. Int. J. Multiscale Com. 10 (2012), pp. 213227.
[6] S. J. Mistry and S. Govindjee. A micro-mechanically based continuum model for strain-induced
crystallization in natural rubber. Int. J. Solids Struct. 51 (2014), pp. 530539.
[7] O. C. Zienkiewicz, R. L. Taylor, and J. Z. Zhu. The Finite Element Method: Its Basis and Funda-
. Elsevier Science, Butterworth-Heinemann, 2013.
mentals, 7th edition
154

Crystallization of semi-crystalline polymers simulated

with a cellular automaton
1
Tobias Daniel Horn ∗
, Hans Wulf1 and Jörn Ihlemann1
1 Chair of Solid Mechanics, Chemnitz University of Technology, Reichenhainer Str. 70, D-09126 Chemnitz, Germany
Abstract. During crystallization of semi-crystalline polymers from melt, structures evolve at different scales.
When crystallized from viscous melts or solutions at moderate undercooling, the polymer forms lamellae by
chain folding at the nanometer scale. These lamellae grow predominantly in radial direction and branch
irregularly by non-crystallographic branching. By this, spherulites are formed at the micrometer scale.
In this contribution, a simulation model based on a concept of isothermal crystallization of semi-crystalline
polymers is derived. To this end, the polymer chains are divided into four different categories concerning their
mobility. For these categories, different rules for diffusion and conversion are defined, which cover the physical
processes the chains experience during crystallization. With these rules, the crystallization of the polymer to
lamellae is modeled. The simulation is realized by using a cellular automaton, which is a natural choice to
model growth processes due to its local character.
Introduction
Crystallization in polymer systems, which transfers the entangled melt into a semi-crystalline state, is a
process of primary importance and has been studied since long time. Over the years various conceptual
models have been proposed [1].
In the multistage model, Strobl [2, 3] categorize the crystallization process into three different phases:
the amorphous phase, the mesomorphic layers and the crystal. In the amorphous phase the polymer
chains form irregular and entangled coils with a high inner mobility. During crystallization the polymer
chains rearrange and form ordered regions called lamellae, where the polymer chains are both aligned
and folded. Here, the chains are fixed due to physical bonds. However, crystallization is not a discrete
process. In front of the lamellae (in the growing direction) the amorphous phase rearrange and a thin
layer with a mesomorphic inner structure forms between the crystal front and the melt. Within this
phase the polymer chains still have an enhanced inner mobility, but the structure of the crystalline phase
is already formed.
1 Simulation Model
The crystallization of semi-crystalline polymers is simulated by using a cellular automaton. Therefore,

at each cell of the automaton the polymer is differentiated into four phases. For this the mobility of the
chains is taking into consideration. In the amorphous phase the polymer chains have a high inner mobility.
In front of the lamellae the polymer chains rearrange, physical bonds are evolved and the mobility of
the chains is restricted. According to the multistage model in this phase the polymer crystallizes under
formation of blocks. Hence, the phase is called blocked phase. In the crystalline phase there are a
maximum of physical bonds and for this the mobility of the polymer chains is limited to a minimum.
Furthermore, due to chain folding there is another phase at the lateral side of the lamellae, where at
least one part of the polymer chain is restricted in his mobility (because the chain end is integrated in
the blocked phase or the crystal phase). Therefore the phase is called the trapped phase. The phases are
visualized in figure 1. As e.g. shown in figure 1 c) a polymer chain can belong to more than one phase.
∗ Corresponding author: tobias.horn@mb.tu-chemnitz.de ORCID iD: 0000-0002-7396-9563
155
Figure 1. Model concept of the four phases in the polymer during crystallization. In the figures the polymer
chains within these phases are visualized. Figure a) shows the amorphous phase, b) the blocked phase,
c) the trapped phase (marked with arrows) next to the crystalline phase, which is also shown in d).
The mobility of the phases is modeled by using Fick’s second law of diffusion (anisotropic and for mul-
ticomponent systems). Due to their low mobility, in the blocked phase and the crystalline phase the
diffusion is suppressed. The remaining dimensionless diffusion coefficients are defined at every cell of the
automaton by using the two parameters D̄a and D̄t

Da Dat D̄a D̄a D̄t
= (1 − ρ) , (1)
Dta Dt D̄a D̄t D̄t
where ρ is the local density.

During crystallization the chains convert from the amorphous phase to the other phases. This is realized
by a Markov chain as shown in figure 2. The related transition matrix is given in equation (2). The
matrix elements are locally defined at every cell of the cellular automaton and depend on the state of the
Moore region around the cell.
 
Maa Mab Mat 0
 Mba Mbb 0 Mbc 
[M ] = 
 Mta
 (2)
0 Mtt 0 
0 0 0 Mcc
Figure 2. Markov chain with all possible conversions in one time step. Pa , Pb , Pt and Pc are the percentages of
the several phases of the polymer at one cell of the cellular automaton.
2 Simulation Results
With these rules for the cellular automaton the crystallization is simulated. Here, the thickness of the
lamellae hL is a model parameter and the growing direction α = 25◦ is set as a constant parameter. For
this simulation with 200 time steps and a cellular automaton with 3600 cells the calculation time was
15s.
In figure 3 the distribution of the crystalline phase is shown at the cells of the cellular automaton. Therein,
the gray vertical and horizontal lines mark the borders of the cells. Grayscale is used to indicate the
156
Figure 3. Distribution of the crystalline phase after 200

time steps. At the grayscale black indicates
Pc = 0 and white Pc = 1. The marked path Figure 4. Distribution of the several phases on the path
is for the evaluation in figure 4 marked in figure 3 after 200 time steps
distribution of the crystalline phase, where black indicates Pc = 0 and white Pc = 1. In figure 4 the
distribution of the several phases in the crystal, at the front of the crystal and in the amorphous phase
is shown. At the front of the crystal, the amorphous phase is converted into the blocked phase and the
trapped phase, as claimed in section 1. In the crystal the amorphous phase is not present, but still the
blocked and trapped phases. They are interpreted as crystal defects.
3 Conclusion
Based on the multistage model of Strobl [2, 3] the crystallization process is modeled by using a cellular
automaton at the scale of the lamellae. The basic principles of the simulation model was presented and
the result of an easy simulation of the crystallization of a single lamellae was shown.
Acknowledgments
The authors gratefully acknowledge the German Research Foundation (Deutsche Forschungsgemeinschaft,
DFG) for supporting this work.
References
[1] Joseph D Menczel. The rigid amorphous fraction in semicrystalline macromolecules. Journal of ther-
mal analysis and calorimetry 106.1 (2011), pp. 7–24.
[2] G Strobl. A thermodynamic multiphase scheme treating polymer crystallization and melting. The
European Physical Journal E 18.3 (2005), pp. 295–309.
[3] Gert R Strobl. The physics of polymers. Springer, 2007.
157
158

Towards thermo-viscoelastic experimental characterisa-

tion and numerical modelling of VHB polymer
Mokarram Hossain∗1 and Zisheng Liao2
1 Zienkiewicz Centre for Computational Engineering, Swansea University, UK
2 School of Civil Engineering and Transportation, South China University of Technology, China
Abstract. In this study, we have presented a wide variety of temperature experiments of a commercially
available VHB polymer ranging from -30◦ C to 80◦ C at various strain rates and stretch levels under homo-
geneous deformation and temperature fields. The study demonstrates a pronounced influence of the tem-
perature field on mechanical responses of the VHB polymer. After a wide range of experiments, we have
proposed a finite-strain thermo-viscoelastic constitutive model where a non-linear evolution law is devised
based on the classical concept of multiplicative decomposition of the deformation gradient. Then, decou-
pled one-dimensional equations are fitted to identify relevant material parameters appearing in the model.
The thermo-viscoelastic model validation shows its excellent capability to predict the experimental results.
1 Introduction
The so-called Very High Bond (VHB in short) becomes an ideal polymer for producing prototypes of
electric field-responsive functional materials, e.g., actuators in soft robotics, stretch sensors in wearable
devices, and energy harvesters from ambient motions. The commercially available polymer comes up
with several thicknesses. For this study, we have selected VHB 4905 thanks to its wide use as a common
material for dielectric elastomers. The acrylic-based polymer is highly deformable, extremely viscoelastic,
and highly sensitive to temperature fluctuations. Hence, in order to understand its mechanical and
electro-mechanical behaviours, extensive experiments need to be conducted to demonstrate temperature
dependencies in addition to rate-dependences. The aims of the current work are twofolds. On the one
hand, a complete thermo-viscoelastic mechanical characterization of VHB 4905 is presented considering
some standard experiments suitable for viscoelastic polymers at large strains, e.g. loading-unloading cyclic
tests, quasi-static tests at very slow strain rates, single-step relaxation tests, and multi-step relaxation
tests under homogeneous thermo-mechanical loads. On the other hand, a thermodynamically consistent
phenomenologically-motivated geometrically non-linear thermo-viscoelastic material model is developed
for VHB polymers. At this stage, all experiments are performed under a homogeneous deformation, both
mechanical and thermal loads. Hence, the three-dimensional thermo-viscoelastic model is decoupled to a
set of one-dimensional equations. After that, all relevant parameters appearing in the model are identified
by a simultaneous optimisation routine. Once all necessary material parameters are identified, the model
fits well with the data that are not included during the identification process, see [1, 2].
2 Experimental study
In order to characterize viscoelastic behaviours of the VHB polymer, at first cyclic experiments are carried
out with regard to different deformations, strain rates, and temperatures. In these experiments, we adopt
three different deformation levels, i.e., 100%, 200%, and 300%, three different strain rates, i.e., 0.1 /s, 0.05
/s and 0.03 /s, and eight different temperature profiles, i.e. 80◦ C, 60◦ C, 40◦ C, 20◦ C, 0◦ C, -20◦ C, -30◦ C
and -40◦ C. Results are presented in Figs 1 and 2. In Fig 1, temperature-dependent results are presented
ranging from -20◦ C to + 60◦ C temperatures while experiments are conducted at a strain rate of 0.1/s
with deformations of 300%. These graphs indicate that at all tested temperatures, a larger deformation
∗ Corresponding author: mokarram.hossain@swansea.ac.uk ORCID iD: ORCID:https://orcid.org/0000-0002-4616-1104
159
0.18
0.1 /s 300% 0.1 /s 300%
-30°C 0.16 10°C

8
-20°C 20°C
-10°C 0.14 40°C

Nominal Stress [MPa]

0°C 60°C
80°C
0.12
6
0.10
0.08
4
0.06
2 0.04
0.02
0 0.00
1.0 1.5 2.0 2.5 3.0 3.5 4.0 1.0 1.5 2.0 2.5 3.0 3.5 4.0
Stretch [1] Stretch [1]
Figure 1. Temperature-dependent stress responses under a wide range of temperature profile where experiments
are conducted at 300% deformation and 0.1 /s strain rate, (Left) below 0◦ C, (Right) above 0◦ C. All
results clearly reveal that the temperature field has a major role in changing mechanical stress responses
of VHB
20°C 20°C
0.4
25%
0.35 25%
100% 100%
200% 0.30 200%

300% 300%
0.3
400% 400%
0.25
500% 500%
600% 600%
700% 0.20 700%
0.2 800% 800%
0.15
0.10
0.1
0.05
0.0 0.00
0 2 4 6 8 10 1 2 3 4 5 6 7 8 9
Time [hr] Stretch [1]
Figure 2. (Left ) Single step relaxation test results from 25% to 800% deformations. (Right) Stress-strain curves
of multi-step tests depicting the amount of stress relaxations during the holding times
results in a bigger hysteresis with a larger residual strain. Moreover, it is clear that the nonlinearity of
the stress-stretch response is getting stronger from a higher to a lower temperature at the same strain.
These results also demonstrate that the VHB 4905 becomes less dissipative with an increasing positive
temperature.
Single-step relaxation experiments are performed on tensile specimens from 100% to 800% deformations
with an internal of 100% deformations. For a better demonstration purpose, the stress history curves are
shifted along the time axis to avoid overlapping of the curves, as is shown in Fig 2(left). From the results,
it can be observed that with the time evolving, the stress attenuates to an equilibrium level. However,
at a larger strain, it takes a longer time to relax. This implies that at a larger deformation level, more
relaxation time is required to obtain an equilibrium stress. Another way to identify the equilibrium stress
of a polymeric material is the multi-step relaxation tests. Similar to the single step relaxation experiments,
the specimen is stretched to a certain strain level and is allowed to relax for a three-hour period before
the specimen is re-loaded to the next strain level instead of being unloaded. In this case, a multi-step
relaxation experiment is carried out at an interval of 100%, 300%, 500% , and 700%. As is shown in
Fig 2(right) the single step relaxation tests and the corresponding steps of the multi-step relaxation tests
share similar relaxation history curves with very close relaxed stresses (equilibrium stresses).
160
3 Numerical modelling
In uniaxial tests, the specimen is elongated only in one direction, i.e., λ1 = λ while the other two
directions are free to move. From the incompressibility condition at fixed temperature, i.e., detF =
detF̄ = detF v = detF Θ = 1, i.e., λ1 λ2 λ3 = 1 and the assumption of symmetry the complementary
principal stretches follow as λ2 = λ3 = λ−1/2 . Therefore, the complete deformation gradient reads
F = {λ; λ−1/2 ; λ−1/2 }. Note that since the elongation is only in one direction, the specimen will contract
in the transversal directions and due to the stress-free boundary conditions, both Piola stresses P2 and
P3 are zero and only the Piola stress P1 needs to be determined. The elastic part of the Piola stress is
obtained:
h i
Θ 1
P e
= + g (Θ) 2a + 8b[2λ−1 + λ2 ]3 + c[1 + 2λ3 ]− 2 λ − λ−2
e
(1)
Θ0
for details, see [1, 3, 4]. Note that since we have observed negligible influence of the temperature on the
equilibrium response during our study, g e (Θ) = 1 − ΘΘ0 is taken here. Now, the viscous part of the stress
for a single Maxwell element is formulated :
X2 X5
v Θ v v,j Θ v
P = + g1 (Θ) P + + g2 (Θ) P v,j
Θ0 j=1
Θ 0 j=3
X2
" #2 " #
2 j
Θ v j λ 2λv λ λjv
= + g1 (Θ) 6c4 + −3 − 2
Θ0 j=1 [λjv ]2 λ [λjv ]2 λ
X5
" #
Θ λ λjv
+ + g2v (Θ) 2cj5 − , (2)
Θ0 j=3 [λjv ]2 λ2
where λ2v is the principal stretch of the internal variable. Similar to the one-dimensional formulation of
the total stress, the first evolution law is derived:
" #2
˙j 4 λ2 2λjv λ2 [λjv ]2
λv = j + −3 − , j = 1, 2 (3)
τ1 [λjv ]2 λ λjv λ
and the one-dimensional form of the second evolution equation is

˙ 4 λ2 [λjv ]2
λjv = j − , j = 3 · · · 5. (4)
3τ2 λjv λ
Now both the scalar-valued differential equations need to be discretized by a suitable integration scheme.
Discretizing by the unconditionally stable implicit Euler-backward integration scheme, the differential
equation for the internal variable λjv for a single Maxwell element yields

4∆t λ2 [λjv ]2
λjv = λj,n v + − ,
3τ2j λjv λ

j j n,j 4∆t λ2 [λjv ]2
f (λv ) = λv − λv − − (5)
3τ2j λjv λ
k
where [•] = [•] (tk ) and ∆t = tn+1 − tn . For the sake of brevity, the index for the current value (tn+1 ) of
the stretch λjv has been dropped in Eqn (5). Eqn (5) is non-linear in terms of λjv . Therefore, a Newton-
type iterative scheme is required to find the current value of λjv that need to be inserted in Eqn (2) to
obtain the updated value of the total Piola stress, P (t) = P e + P v (t).
4 Model validation
For the identification of hyperelastic material parameters appearing in a model, three main procedures
have been discussed in the literature, see [2]. The first option is to perform single-step relaxation experi-
ments at various strain levels. When a test specimen is held at a certain strain, the stress will gradually
161
0.3 800% -20°C, 300%

Experi 0.05 /s 1.6
Exper 0.05 /s
Simul 0.05 /s Simul 0.05 /s
0.25 1.4
Experi 0.03 /s Exper 0.03 /s

Simul 0.03 /s Simul 0.03 /s
1.2
Experi 0.01 /s
0.2
Simul 0.01 /s 1
0.15 0.8
0.6
0.1
0.4
0.05
0.2
0 0
1 2 3 4 5 6 7 8 9 1 1.5 2 2.5 3 3.5 4
Stretch [1] Stretch [1]
Figure 3. Viscoelastic parameter identification: Three different strain rates are used to identify viscous parameters
while hyperelastic parameters identified earlier are kept frozen. Total five Maxwell elements, each
having two material parameters, are well enough to capture the strain rate-dependent data up to 800%
deformations. (Right) Model performance is validated by other strain rate dependent tests
reduce and will asymptotically reach an equilibrium stress state, which can be considered as the relaxed
and basic elastic stress at the corresponding strain. The next step is to identify the viscoelastic param-
eters. With the hyperelastic parameter, the equilibrium stress can be calculated, and the over-stress
modulus and over-stress relaxation spectrum are yet to be identified in order to obtain the viscous over-
stress. Since viscoelasticity is the strain rate dependent behavior, we identify the parameters using three
sets of data at 800% strain and at different strain rates, which are 0.01 /s, 0.03 /s, and 0.05 /s, simul-
taneously. Note that during the identification process, we find that five sets of Maxwell elements can
capture the viscoelastic behavior with a fair accuracy while more elements could not improve the results
significantly, cf Fig 3 (left). During the viscoelastic parameter identification, hyperelastic parameters are
kept fixed. With the identified parameters, both strain rate dependences and large strain nonlinearities
are captured successfully, see Fig 3(right).
5 Conclusion
In this contribution, a detail thermo-viscoelastic experimental characterizations of a widely-used VHB

polymer is presented. All experiments were performed in a purposively-built temperature chamber that
can operate under a wide span of temperature ranging from -50◦ to 150◦ C. In order to quantify viscoelastic
behaviours of the polymer, all classical experimentations, i.e., quasi-static tests at very low strain rates,
single-step and multi-step relaxation tests, loading-unloading cyclic tests were conducted at different
strain rates and at different temperature profiles. Exploring the experimental data obtained in the current
study, a thermodynamically consistent phenomenologically motivated thermo-viscoelastic material model
at finite strain is proposed. Once, a complete set of viscoelastic parameters are identified, we use them
to validate other sets of data that are not utilised in any identification process. All validation examples
show an excellent agreement between model and experimental data.
References
[1] Z. Liao, M. Hossain, X. Yao, M. Mehnert, P. Steinmann, On thermo-viscoelastic experimental char-

acterisations and numerical modelling of VHB polymer, In Review, 2019
[2] M. Hossain, D. K. Vu, P. Steinmann, Experimental study and numerical modelling of VHB 4910
polymer, Computational Materials Science, 59:65-74, 2012
[3] P. Steinmann, M. Hossain and G. Possart, Hyperelastic models for rubber-like materials: Consistent
tangent operators and suitability of Treloar’s data, Arch. of Appl. Mech., 82(9):1183-1217, 2012
[4] M. Hossain, P. Steinmann, More hyperelastic models for rubber-like materials: Consistent tangent
operator and comparative study, Journal of Mechanical Behaviour of Materials, 22(1-2):27-50, 2013
162

A Macroscopic Model for Magnetorheological Elastomers

based on Microscopic Simulations
Karl A. Kalina ∗1
, Philipp Metsch1 , Jörg Brummmund1 and Markus Kästner1,2
1 Chair of Computational and Experimental Solid Mechanics, TU Dresden, 01062 Dresden, Germany.
2 Dresden Center for Computational Materials Science (DCMS), TU Dresden, 01062 Dresden, Germany.
Abstract. Herein, the parametrization of a macro-model for magneto-rheological elastomers based on mi-
croscopic simulations is presented. Within a computational homogenization, the effective response of the
composite system is calculated and the data are used for parameter identification. The merit of this strategy
is the adjustment of the model independent of any macroscopic sample geometry. With the developed model,
the magnetostrictive behavior of a macroscopic sample is simulated.
Introduction
Magnetorheological elastomers (MREs) are a class of active composites which consist of a polymer matrix
filled with micron-sized magnetizable particles. Thus, the effective stiffness or the shape of a specimen
can be controlled with an external magnetic field.
Continuum based strategies regarding the modeling of MREs can be divided into microscopic [5, 6, 9,
10] and macroscopic [1, 2, 3, 4] approaches. In the former, the heterogeneous microstructure is explicitly
resolved and it is possible to predict the effective behavior of the composite. However, the disadvantage of
microscopic continuum models is the computationally expensive numerical solution. In contrast to that,
macroscopic models, which consider the MRE as a homogeneous continuum, enable the efficient simulation
of real structures. Since such models are phenomenologically motivated, the model parameters have to
be determined which is done by using experimental results so far, e. g. [2, 3]. Due to the inhomogeneous
fields inside the MRE sample [7, 8], this procedure however requires several simplifying assumptions.
In this contribution a macroscopic model for MREs is identified from data generated by a microscopic
model combined with a computational homogenization approach. This strategy enables to identify the
parameters independent of any sample geometry.
The presented approach is based on a general continuum formulation of the coupled magneto-mechanical
boundary value problem [5, 6, 7, 10] which is valid on the micro- as well as the macroscale. To link these
two scales, a suitable homogenization scheme is applied [9, 10].
1.1 Microscopic Simulations
For the description of the macroscopic MRE-behavior in the isotropic case, a random microstructure
has to be taken into account. In order to ensure a statistical representation, the number of embedded
particles is chosen to N = 300, where representative volume elements (RVEs) with the characteristic
values φ = {10, 15, 20, 25, 30, 35, 40} % are generated. To estimate mean values as well as confidence
∗ Corresponding author: karl.kalina@tu-dresden.de
163
M1 M2 M3 MM1 MM2
H H
MM3 MM4 MM5 MM6 MM7
H H
H H H
Figure 1. Schematic illustration of the simulated purely mechanical load cases M1 – M3 and coupled magneto-
mechanical load cases MM1 – MM7 for an RVE with φ = 30 %. The effective deformation and the
magnetic field are applied simultaneously within the homogenization. The maximum applied effective
stretch and shear deformation are chosen to λ = 1.15 and γ = 0.15 for the mechanical load cases. In
the coupled cases, the values |H| = 230 kA m−1 , λ = 1.1 and γ = 0.1 are applied.
intervals for the effective magneto-mechanical response, five different microstructures are considered for
each particle-volume fraction.
All RVEs are analyzed by means of three pure mechanical load cases (M1 – M3) and seven coupled
magneto-mechanical load cases (MM1 – MM7) in the linear magnetic regime, see Fig. 1. Within the
microscopic simulations, the constitutive models presented in [5, 10] are used. All performed simulations
are realized for the plane strain case.
1.2 Macroscopic Model
In order to enable a separate description of deviatoric and volumetric parts of the MRE behavior, the
deformation gradient F is divided into isochoric F̄ = J −1/3 F and volumetric contributions according
to the Flory-split, where J = det F denotes the Jacobi determinant. Due to the assumed isotropy, the
MREs effective response can be described in terms of the six invariants
1 2 −1
I¯1 = tr C̄ , I¯2 = (tr2 C̄ − tr C̄ ) , I3 = J 2 , I4 = |H|2 , I¯5 = H · C̄ · H and I¯6 = H · C̄ · H . (1)
2
Therein, C and H denote the right Cauchy-Green deformation tensor and the Lagrangian magnetic field,
respectively. For a separate discussion, the amended free energy function
Ω = Ωmech (C) + Ωcoup (C, H) + Ωmag (H) + Ωfree (C, H) (2)
is split into mechanic, coupling, magnetic and free space part, where the latter is the Lagrangian coun-
terpart of the magnetic free field energy and is independent of any material properties. With that, the
ansatz
1h i h i µ β
0 µ0
Ω = C(I¯1 − 3) + K(I3 − 1) + µ0 α1 I¯5 + α2 I¯6 + α3 (I3 − 1)I¯6 + I4 − JC −1 : (H ⊗ H) (3)
2 2 2
is chosen for the macro-model, where µ0 is the permeability of vacuum.
1.3 Parameter Identification
To ensure a high accuracy of the model, it should describe the effective magnetization m as well as the
total and mechanical stress tensors σ tot and σ. The material parameter sets κ = {κmech , κcoup , κmag }
164
(a) (b) (c)

4 |m | / kAm−1 C11
0 1.11
152
3
σ / kPa
ε/%
−1 h∞
2
−2 1
0 100 200 10 15 20 25 30 35 40 x2 149 0.952
|h | / kAm−1 φ/%
x3 x1
σ11 σ12 σ21 σ22 model |b | = const
Figure 2. Simulation of MREs: (a) average of the homogenized mechanical stress with 95 % error bound and fitted
model for load case MM1 with φ = 20 %, (b) overall magnetostrictive strain ε = ∆l/l0 in dependence of
φ and (c) local magnetization |m| and deformation field C11 within a deformed circular MRE sample
with φ = 30 %. The undeformed configuration is marked with dashed lines and the depiction of the
deformed configuration is scaled with the factor five.
are determined within a stepwise algorithm which results in three linear optimization schemes. Finally,
to achieve a good approximation of all considered fields, the inherently nonlinear step
(
XN 2
coup mag
{κ ,κ } = arg coup
minmag Wiσ σ(F i , H i , κ) − σ i ...
κ κ
i=1
)
2

...+Wim m(F i , H i , κcoup , κmag ) − mi (4)
is realized, where the parameters determined in the previous steps are used as initial values. Wim and Wiσ
denote weighting factors for the single error sums which are determined as the inverse of the maximum
value of σkl and mk for each load case, respectively.
2 Results
The fitting procedure is applied to the homogenized data of the load cases M1–M3 as well as MM1–MM3,
where Fig. 1(a) exemplarily depicts σ for MM1 with a volume fraction of φ = 20 %. The model deviations
of the load cases which where not used within the parameter identification procedure are in a similar
range as the fitted ones. Thus, the quality of the developed model could be validated.
Finally a circular MRE sample within a surrounding free space is simulated, where special attention is
drawn to the magnetostriction [8, 9], i. e. the elongation or contraction in an external field h∞ . As
shown in Fig. 1(c), the induced magnetization within the sample is almost homogeneous, whereas the
deformation field is strongly inhomogeneous. In case of a rectangular shaped sample, this effect intensifies
[7, 8]. This demonstrates again, that the parameter identification for macroscopic MRE models from
experimental results requires several assumptions in the fitting process.
If the underlying particle volume fraction of the MRE is varied, a degressive curve for the overall mag-
netostrictive strain ε = ∆l/l0 is observed if the applied local induction b is set constant, see Fig. 1(b).
Altogether, the predicted elongation of the sample in the direction of the external field is in qualitative
accordance with experiments for isotropic MREs.
Acknowledgments
The present study is funded by the German Research Foundation, Priority Programme 1681 grant KA
3309/2-3. This support is gratefully acknowledged.
165
References
[1] B.E. Abali and A.F. Queiruga. Theory and computation of electromagnetic fields and thermome-
chanical structure interaction for systems undergoing large deformations. Journal of Computational
Physics 394 (2019), pp. 200 –231.
[2] R. Bustamante, A. Dorfmann, and R.W. Ogden. Numerical solution of finite geometry boundary-
value problems in nonlinear magnetoelasticity. International Journal of Solids and Structures 48.6
(2011), pp. 874 –883.
[3] K. Danas, S. V. Kankanala, and N. Triantafyllidis. Experiments and modeling of iron-particle-
filled magnetorheological elastomers. Journal of the Mechanics and Physics of Solids 60.1 (2012),
pp. 120–138.
[4] K. Haldar, B. Kiefer, and A. Menzel. Finite element simulation of rate-dependent magneto-active
polymer response. Smart Materials and Structures 25.10 (2016), p. 104003.
[5] K. A. Kalina, P. Metsch, and M. Kästner. Microscale modeling and simulation of magnetorheological
elastomers at finite strains: A study on the influence of mechanical preloads. International Journal
of Solids and Structures 102–103 (2016), pp. 286 –296.
[6] K. A. Kalina, J. Brummund, Metsch P., M. Kästner, D. Yu. Borin, J. M. Linke, and S. Odenbach.
Modeling of magnetic hystereses in soft MREs filled with NdFeB particles. Smart Materials and
Structures 26 (2017), pp. 105019 –105031.
[7] K. A. Kalina, P. Metsch, J. Brummund, and M. Kästner. Development of a Macro-Model for
Magnetorheological Elastomers based on Microscopic Simulations. International Journal of Soldis
and Structures (submitted 2019).
[8] M.-A. Keip and M. Rambausek. Computational and analytical investigations of shape effects in the
experimental characterization of magnetorheological elastomers. International Journal of Solids and
Structures 121 (2017), pp. 1 –20.
[9] P. Metsch, K. A. Kalina, Christian Spieler, and M. Kästner. A numerical study on magnetostrictive
phenomena in magnetorheological elastomers. Computational Materials Science 124 (2016), pp. 364
–374.
[10] P. Metsch, K. A. Kalina, J. Brummund, and M. Kästner. Two- and three-dimensional modeling
approaches in magneto-mechanics: a quantitative comparison. Archive of Applied Mechanics 89
(2018), pp. 47–62.
166

Experimental Characterization, Modeling and FE-Si-

mulation of Curing Phenomena in an Elastomeric Ad-
hesive
Ralf Landgraf
∗ 1 and Jörn Ihlemann
1
1 Chair of Solid Mechanics, Chemnitz University of Technology, Chemnitz, Germany
Abstract.
This contribution deals with the curing process of a PU based adhesive. The basic experimental methods and
corresponding modeling approaches are introduced. Moreover, an application of the resulting curing model to
the simulation of the production process of piezo-metal compounds by nite element simulations is shown.
Introduction
The crosslinking (or curing) process of a PU based adhesive is considered. The raw material is a one-
component paste with enclosed encapsulated isocyanate. The curing process is initiated by raising the
temperature above 90◦ C . Once the curing process has started, the material cures to a solid within few
minutes and the nal material exhibits rubber-like material behavior.
Aiming for the simulation of the material behavior, both experimental characterization methods and phe-
nomenological modeling approaches were applied, which is shown in Section 1. Moreover, an exemplifying
nite element simulation provided in Section 2 will highlight the capability of the modeling approach.
1 Experimental characterization and material modeling
1.1 Progress of the chemical curing process
Dierential scanning calorimetry (DSC) measurements were applied to characterize the chemical crosslink-
ing process (see [4]). Thereby, the rate of the specic reaction enthalpy ḣ was measured during isothermal
curing experiments at temperatures θiso . Additionally, the ultimate specic heat reaction hult , i.e. the
maximum reaction enthalpy that the material releases within a complete curing reaction, was experi-
mentally determined. These values were employed to dene the degree of cure q ∈ [0, 1] by its rate
ḣ(θiso )
q̇ = (1)
hult
Next, a phenomenological curing model was formulated. It is based on the Kamal and Sourour autocat-
alytic reaction model [3] in combination with Fournier's diusion factor fD (q, qend (θ)) [2]:
β k1
q̇ = K1 (θ) + K2 (θ) q α 1 − q fD (q, qend (θ)) , qend (θ) = − k3 . (2)
1 + exp − k2 (θ − θref )
The factors K1 (θ) and K2 (θ) are temperature-dependent functions that show Arrhenius behavior. For
the maximum attainable degree of cure at certain temperatures qend (θ), the ansatz (2)2 was chosen. The
curing model includes the material parameters α, β , k1 , k3 , k3 as well as further parameters inside the
Arrhenius functions and Fournier's diusion factor. A summary of the identication procedure, where the
∗ Corresponding author: ralf.landgraf@mb.tu-chemnitz.de ORCID iD: 0000-0001-7715-5674
167
model (2) is adapted to the experimental data (1), and a list of the identied parameters can be found in
[4]. The resulting behavior of the model is exemplied in Fig. 1 (left) for dierent curing temperatures.
Therein, it can be seen that curing only takes place at temperatures above 90◦ C.
1.2 Heat expansion and chemical shrinkage
Volume changes due to varying temperatures and chemical crosslinking were analyzed by a test setup
based on Archimedes' principle (for details, see [8]). However, instead of using long term measurements
(as applied in [8]), only specic states were analyzed for the current material due to its specic curing
90◦ C). The tested sequence includes measurements
characteristics (very fast curing at temperatures above
◦ ◦ ◦
of the uncured material (q = 0) at temperatures of 25 C, 50 C and 75 C. Next, the material was cured at
◦
95 C and the mass density of the fully cured material was measured. Finally, a sequence of decreasing
test temperatures was applied to identify the mass density of the fully cured materials (q = 1). The
resulting mass densities of two dierent specimens are depicted in Fig. 1 (right). Moreover, the results
of a fully cured test specimen (experiment 3) are included as well.
Fig. 1 (right) reveals decreasing mass densities for increasing temperatures (heat expansion). Moreover,
the mass density decreases during the phase change from liquid to solid state, which is in contrast to the
well-known chemical shrinkage during curing. The increasing volume during the curing process could be
attributed to the formation of pores. The temperature and curing dependent mass density is modeled as
follows:
%e h i
%(θ, q) = , e + βq q , αθ = 4.21 · 104 K−1 , βq = 6.06 · 10−2 .
ϕ(θ, q) = exp αθ (θ − θ) (3)
ϕ(θ, q)
kg
Therein, %e = 1515 m 3 is the reference mass density, which was measured at the reference temperature
θe = 25 C and the reference degree of

◦
cure qe = 0. The function ϕ(θ, q) is the thermochemical volume
ratio, for which a phenomenological ansatz was chosen. The ansatz includes the heat expansion coecient
αθ and the chemical expansion coecient βq . The corresponding parameter values for the investigated
adhesive are given in Eq. (3).
1 1600
model
degree of cure q [−]
0.8 1550 q=0

density ρ [kg/m3]
110°C experiments 1−3

100°C
0.6 95°C 1500
92°C
0.4 90°C 1450
85°C
0.2 1400
q=1
0 1350
0 5 10 15 20 25 30 20 30 40 50 60 70 80 90 100
time t [min] temperature θ [°C]
Figure 1. Simulation of the evolution of the degree of cure q for dierent constant temperatures (left), and results
of mass density measurements (right).
1.3 Mechanical behavior
In a rst step, the mechanical behavior was characterized by uniaxial tension tests on fully cured speci-
mens. The resulting stress-strain curves of loading/unloading cycles with increasing amplitudes revealed
typical phenomena of rubber-like materials (see Fig. 2, left). Because of this, the Model of Rubber
Phenomenology (MORPH, see [1]) was adapted to the experimental data by the help of a parameter
identication procedure (see [4]). Thereby, the eight included material parameters pe1 − pe8 were identied
with respect to a reference temperature θe = 25◦ C and a reference degree of cure qe = 1. Fig. 2 (left)
depicts the comparison of a tensile test and the corresponding simulation of the mechanical behavior with
the MORPH model. Therein, a good accordance between experiments and simulation can be observed.
168
Next, the MORPH model was extended by temperature and degree of cure eects. Thereby, the reference
material parameters were replaced by the parameter functions:
pi (θ, q) = pei · fq (q) · fθ (θ) , i = 1...8 , (4)
where fq (q) is a factor depending on the degree of cure. This function was adapted to measurements
of increasing storage moduli, which were obtained by rheometer tests on the adhesive during its curing
process. In addition, fθ (θ) is a temperature dependent function that was adapted to temperature ramp
experiments on the cured material within dynamical mechanical analyses (DMA). An example of the
evolution of the stress-strain curves within a curing process is given in Fig. 2 (right). Therein, the
increasing stiness and the inelastic material behavior can be observed. More details on the modeling
approach can be found in [4].
12 20
1st PK stress Txx [MPa]
1st PK stress Txx [MPa]

experiment
10 MOPRH model 15
8 q=0.98
10
6
5 q=0.5
4
2 0
q=0.0
0 −5
1 1.2 1.4 1.6 1.8 2 1 1.2 1.4 1.6 1.8 2
stretch λx [−] stretch λx [−]
Figure 2. MORPH model: adaption to uniaxial tension tests (left) and simulation of evolving stress-strain behavior
within a curing process (right).
2 Finite element simulations
The model outlined in Section 1 was implemented into the commercial nite element software ANSYS
(see [6] and [5]) to simulate the production process of smart sheet metal compounds with included sensor
and actuator functionality. Such a piezo-metal compound (PMC) consists of two metal sheets that are
connected by an initially liquid adhesive layer. Moreover, a sensor/actuator (like macro-ber-composites)
is embedded inside the adhesive layer (see Fig. 3, left). The production of a 3D-curved PMC consists of
three dierent steps according to Fig. 3. First, a local heating/cooling step causes the outer regions of
the adhesive layer to cure while the inner adhesive region remains in liquid state. Next, the pre-cured
compound is formed to its nal state. The liquid adhesive core thereby protects the sensor/actuator and
the cured outer adhesive region prevents the liquid adhesive core to leak. In the nal step, the adhesive
core of the formed PMC is cured to obtain a complete closure of the compound.
Figure 3. Structure and three basic steps for the production of PMCs.
The steps for the production of PMCs were analyzed by nite element simulations. First, the local heat-
ing/cooling step on at PMCs was investigated by pure thermal FE simulations. Thereby, heating/cooling
was induced by convection boundary conditions on the metal sheets. The arising temperature distribu-
tion within the adhesive layer results in local curing (see Fig. 4, left). More precisely, only regions are
cured, where the temperature rises above 90◦ C, i.e. in the outer regions near the heating units. Instead,
the adhesive core remains liquid (q = 0). More results on this production step are reported in [5].
Next, the geometry and the material properties of the cured adhesive were passed to project partners, who
†
simulated the forming of the PMC to a 3D-curved shape. Fig. 4 (middle) depicts the von Mises equivalent
† Simulations were conducted by S. Hensel (SFB/TR39, subproject B1) using the FE software LS-DYNA.
169
Figure 4. Finite element simulation of the production process of PMCs: local heating/cooling (left), forming
(middle), and nal curing (right).
stress eld in all solid parts (metal sheets, sensor, and cured adhesive). It can be seen that only low stresses
act on the sensor/actuator due to the oating support of the liquid adhesive. The nal geometry was then
passed back to simulate the remaining curing process. Due to the complete and homogeneous heating of
the PMC, mechanical nite element simulations with prescribed temperatures were applied to investigate
the eect of volume changes and increasing stiness. The results are exemplied by Fig. 4 (right), where
vertical displacements are depicted that appear due to the curing process. It can be observed that the
previously liquid core increases its volume due to the curing process.
Summary
A combined approach for the experimental characterization and phenomenological modeling of curing
processes in a PU based adhesive has been presented. Moreover, an exemplifying application of the
curing model within nite element simulations has been given. The general approach can also be applied
to dierent polymers and technical processes, see for example [6] and [7].
Acknowledgments
This research was funded by the German Research Foundation (DFG) within the Collaborative Research
Center SFB/TR39 PT-PIESA.
References
[1] D. Besdo and J. Ihlemann. A phenomenological constitutive model for rubberlike materials and its
numerical applications. International Journal of plasticity 19.7 (2003), pp. 10191036.
[2] J. Fournier, G. Williams, C. Duch, and G. A. Aldridge. Changes in molecular dynamics during
bulk polymerization of an epoxide-amine system as studied by dielectric relaxation spectroscopy.
Macromolecules 29.22 (1996), pp. 70977107.
[3] M. R. Kamal and S. Sourour. Kinetics and thermal characterization of thermoset cure. Polymer
Engineering & Science 13.1 (1973), pp. 5964.
[4] R. Landgraf and J. Ihlemann. Application and extension of the MORPH model to represent curing
phenomena in a PU based adhesive. In: Constitutive Models for Rubber X. Ed. by A. Lion and M.
Johlitz. London: Taylor & Francis Group, 2017, pp. 137143.
[5] R. Landgraf and J. Ihlemann. Thermally Controlled Adhesive Curing during the Production of
Piezo-Metal-Compounds: Finite Element Modeling and Analyses. Advanced Engineering Materials
20 (2018). doi: doi:10.1002/adem.201800411.
[6] R. Landgraf, M. Rudolph, R. Scherzer, and J. Ihlemann. Modelling and simulation of adhesive curing
processes in bonded piezo metal composites. Computational Mechanics 54.2 (2014), pp. 547565.
[7] R. Landgraf, J. Ihlemann, S. Kolmeder, A. Lion, H. Lebsack, and C. Kober. Modelling and simulation
of acrylic bone cement injection and curing within the framework of vertebroplasty. ZAMM-Journal
of Applied Mathematics and Mechanics 95.12 (2015), pp. 15301547.
[8] M. Rudolph, M. Stockmann, R. Landgraf, and J. Ihlemann. Experimental determination and mod-
elling of volume shrinkage in curing thermosets. arXiv Preprint arXiv:404.0310 (2014).
170

Modeling the Chemo-Thermo-Mechanical Material Be-

havior of an Epoxy Resin System
Chris Leistner ∗1
and Stefan Hartmann1
1 Institute of Applied Mechanics, Clausthal University of Technology, Adolph-Roemer-Straße 2a, 38678 Clausthal-Zellerfeld, Ger-
many
Abstract. Epoxy resin systems frequently are applied in fiber composite materials for lightweight construc-
tions. During the production process, the epoxy resin system must be cured by an exothermic chemical
reaction, which leads to an inhomogeneous temperature distribution in the part and further affects the curing
kinetics. Especially for thick-walled structures, combustion might occur during curing. Furthermore, chemical
and thermal shrinkage yield residual stresses, warpage or cracks. The goal is to describe these properties
of the material by a material model. The rate-type constitutive equations for the mechanical and the cur-
ing part are implemented into an implicit finite element program, where high-order and time-adaptive time
integration are required to predict precisely the maximum temperature and residual stresses within a part.
Introduction
In this contribution, the thermo-mechanical properties of a pure epoxy resin system are experimentally
investigated and modeled. Regarding further information of the chemo-thermal investigation of the
particular epoxy resin system, we refer to [3]. In [4] the volumetric thermal and chemical deformation
according to the curing reaction is investigated. Such deformation causes mechanical strains and residual
stresses, which are constituted within a viscoplasticity model. The viscous effects are based on a Prony
series of Maxwell elements combined with a hysteretic part using an endochronic model. Of course,
the temperature- and cure-dependent quantities required in the heat equation are provided as well. A
particular focus lies on the challenging parameter identification regarding a unique set of parameters.
Apart from experimental investigations and modeling, the numerical treatment is discussed.
1 Experiments
Tensile tests with dogbone-like specimens of the cured epoxy resin system were carried out as part of
the material investigation. Strain-controlled long-term relaxation tests and multi-step relaxation tests at
four different temperatures serve to separate the state of equilibrium. The basic procedure is similar to
the investigations in [2]. Fig. 1a clearly shows that the material has pronounced relaxation times. In
order to reduce the testing time, the holding time of the multi-step relaxation tests is set to 6 h. Since
the material cannot totally relax within short time, the equilibrium point must be estimated, where we
choose an exponential function using the courses of the long-term relaxation tests, see Fig. 1b.
2 Material Modeling
The rheological model depicted in Fig. 2 serves to represent the material behavior. The strains are
additively decomposed into a thermo-chemical and a mechanical part, where the mechanical part consists
of elastic and a viscous strains:
E = EΘc + EM and EM = Eei + Evi (1)
∗ Corresponding author: chris.leistner@tu-clausthal.de
171
1
60
0.9
0.8 50
0.7
40
σ in MPa
σ in MPa
0.6
30
0.5
0.4 20
0.3
−10 ◦C 10 −10 ◦C
0.2 20 ◦C 20 ◦C
55 ◦C 0 55 ◦C
0.1
85 ◦C 85 ◦C
0 −10
0 2 4 6 8 10 12 14 0 1 2 3 4 5
t in h ε in %
(a) Long time relaxation (b) Multi-step relaxation (the markers define estimated
equilibrium points of relaxation)
Figure 1. Tensile tests at four different temperatures
Keq , Geq
Teeq
beq , aeq
T Theq
Govi ηovi
Tovi
Eei Evi
EΘc EM
Figure 2. Rheological model
Furthermore, the additive decomposition of the stress state

T = Teeq + Theq + Tovi (2)
is proposed as depicted in Fig. 2, where Tovi indicates a sum of Maxwell elements. The stress can
be represented formally as a split into spherical and deviatoric quantities, whereby the spherical stress
consists of a purely elastic equilibrium stress Teeq and the deviatoric stress is decomposed into
D
TD = Teeq D + Theq + Tovi D . (3)
Thus, the total stress power reads
D

T · Ė = −pI · Ė + TD · Ė = −pε̇vol + Teeq D + Theq · ĖM + Tovi D · Ėe + Ėv + TD · ĖΘc . (4)
Due to advantages in constitutive modeling of caloric processes and volumetric expansions, we define a
hybrid free energy, which was introduced in [5]
1
ϕ = ψ + pεvol . (5)
ρ
Under consideration of the hybrid free energy, the Clausius-Duhem inequality reads
1 1 1
δ = −ϕ̇ − εvol ṗ + TD · Ė − sΘ̇ − ~q · ~g ≥ 0 with ~g = grad Θ, (6)
ρ ρ ρΘ
172
where the hybrid free energy can generally be defined by introducing the primary variables
ϕ(t) = ϕ̃(p, EM , Ee , Y, Θ, c) = ϕ̂(p, Θ, c) + ϕ̄(EM , Ee , Y, Θ, c). (7)
Apart from the stress power Eq. (4), the general time derivative of the hybrid free energy Eq. (7) can
be inserted into the Clausius-Duhem inequality, which must be fulfilled for arbitrary values of the time
derivatives of the primary variables in order to satisfy thermodynamic consistency

∂ ϕ̃ ∂ ϕ̃ D ∂ ϕ̃
ρδ = εvol − ρ ṗ + Teeq D − ρ · ĖM + Theq · ĖM − ρ · Ẏ
∂p ∂EM ∂Y
(8)
∂ ϕ̃ ∂ ϕ̃ ∂ ϕ̃ 1
+ TD ov − ρ · Ėe + TD
ov · Ė v + T D
· ĖΘc − ρ ċ − ρ s + Θ̇ − ~
q · ~
g ≥ 0.
∂Ee ∂c ∂Θ Θ
Evaluation of Eq. (8) leads to the constitutive equations listed in Tab. 1. The corresponding material
Table 1. Constitutive equations

Volumetric strains
1
εvol = 3 JΘc 1/3 − 1 − p with JΘc = 1 + αΘ ϑ − αc c − αΘc ϑc (9)
Keq
Elastic equilibrium stress (spherical)
Teeq V = Keq (tr EM )I (10)
Elastic equilibrium stress (deviatoric)

q
αeq1 M1 (Θ)
Teeq D = 2Geq (sM , Θ)ED
M with Geq (sM , Θ) = , sM = ED D
M · EM (11)
αeq2 M1 (Θ) + sM
Hysteretic equilibrium stress (deviatoric)

q
h D D D D D
Ṫeq = aeq M2 (Θ)ĖM − beq ŝ˙ M Theq , ŝ˙ M = ĖM · ĖM (12)
Overstress (deviatoric)
D 2Govi
Ṫovi D = 2Govi ĖM − Tovi D (13)
ηovi
Curing kinetics
n
ċ = fk (Θ, c)fd (Θ, c) with fk (Θ, c) = k(Θ) g + (1 − g) c − c2
1 (14)
and fd (Θ, c) = (1 − tanh (b (ΘG (c) − Θ)))
2
Temperature shift functions

1 Θ − pΘ1 1 ΘG (c) − Θ
M1 (Θ) = 1 − tanh and M2 (Θ) = 1 − tanh (15)
2 pΘ2 2 pΘ3
parameters are identified by step-wise procedures in order to obtain a unique set of parameters.
3 Simulation
As a numerical example a plate with a hole of pure epoxy resin system is considered, which cures within
a cavity at a fixed outer temperature Θs . The dimensions and boundary conditions are depicted in
Fig. 3a. Quadratic hexahedral elements are applied with a concentration at the outer surfaces. In order
173
(a) Geometry, mesh and boundary conditions (b) Displacement
(c) Temperature (d) Degree of cure
Figure 3. Simulation example at t = 350 s
to solve the illustrated problem, we apply diagonally implicit Runge-Kutta (DIRK) methods extended
by time-adaptivity based on [1] and the works cited therein. The resulting distributions of displacement,
temperature and degree of cure at t = 350 s are represented in Figs. 3b to 3d.
References
[1] P. Ellsiepen and S. Hartmann. Remarks on the interpretation of current non-linear finite element
analyses as differential-algebraic equations. International Journal for Numerical Methods in Engi-
neering 51 (2001), pp. 679–707.
[2] S. Hartmann. A thermomechanically consistent constitutive model for polyoxymethylene. Archive of
Applied Mechanics 76.5-6 (2006), pp. 349–366.
[3] C. Leistner, S. Hartmann, D. Abliz, and G. Ziegmann. Modeling and simulation of the curing process
of epoxy resins using finite elements. accepted for publication, Continuum Mechanics and Thermo-
dynamics (2018), pp. 1–24.
[4] C. Leistner, S. Hartmann, J. Wittrock, and K. Bode. Shrinkage behavior of Araldite epoxy resin
using Archimedes’ principle. Polymer Testing 67 (2018), pp. 409–416.
[5] A. Lion, B. Dippel, and C. Liebl. Thermomechanical material modelling based on a hybrid free
energy density depending on pressure, isochoric deformation and temperature. International Journal
of Solids and Structures 51.3 (2014), pp. 729–739.
174

Modelling and simulation approaches for the numerical

analysis of UV curing polymers used in additive manu-
facturing
T. Rehbein ∗ 1
, A. Lion1 and M. Johlitz1
1 Instute of Mechanics, Faculty of Aerospace Engineering, Bundeswehr University Munich
Abstract. This contribution focuses on the experimental characterisation and modelling of the crosslinking
of UV curing polymers used in additive manufacturing processes such as digital light processing and stere-
olithography. First photocalorimetric measurements with varying temperature and light intensity are shown.
From the exothermic specific heat flows measured during the crosslinking reaction, the degree of cure can be
determined for each experimental scenario. It is shown that the test temperature and light intensity have a sig-
nificant influence on the crosslinking reaction. A first modelling approach for the description of the crosslinking
reaction is presented. Moreover, parameter identification and validation of the proposed model are conducted.
Introduction
Additive manufacturing (AM) is an innovative technology to create three dimensional objects with com-
plex geometry. Although the technology has been used in industrial applications for decades, reliable and
experimentally validated models for the design of printed components are still missing. Especially the
material modelling and finite element analysis of UV curing polymers (so-called photopolymers) used in
processes such as digital light processing and stereolithography are challenging tasks. For this purpose,
the crosslinking reaction in the additive manufacturing process must be described correctly by a phe-
nomenological model and adapted to experimental data.
The crosslinking of photopolymers is initiated by the absorption of a light source. Besides that, the
process depends strongly on the ambient temperature: a higher ambient temperature leads to faster
crosslinking whereby the same applies to an increasing light intensity. Calorimetric measurements with
the aid of an additional light source have proved useful for the experimental investigation of the crosslink-
ing reaction of photopolymers [2].
Since the degree of cure directly influences the mechanical, thermal and chemical properties of photopoly-
mers, a validated material model for the degree of cure is essential for finite element analysis of additive
manufacturing processes.
Experimental characterisation
In order to understand and investigate the crosslinking process of commercial photopolymers used in
additive manufacturing processes, differential photocalorimetry (DPC) measurements are conducted using
the TA Instruments DSC Q2000 with an additional light source (OmniCure R
S2000). The light source
contains a high pressure 200 W mercury lamp with an spectral output of 320 to 500 nm.
To investigate the crosslinking reaction, the DSC measures the specific heat flow ḣ between a sample in
a pan and an empty reference pan during the crosslinking reaction. The following relationship is used to
convert the specific heat flow into the degree of cure:
Rt
ḣ t̃ dt̃
0
q (t) = (1)
htot
∗ Corresponding author: rehbein@unibw.de
175
Due to the normalisation to the maximum specific heat htot during the crosslinking reaction, the degree
of cure assumes values between 0 and 1.
The commercial photopolymer LOCTITE R
3D 3830 was used for the photocalorimetric measurements.
It is an acrylic compound for prototype development with low elongation at break, high tensile strength
and low chemical shrinkage. During the photocalorimetry measurement,
following steps are performed
subsequently for the determination of the specific heat flow ḣ t̃ in eq. (1):
1. Equilibration of the measurement cell at a specific temperature (−10◦ C. . . 50◦ C).
2. Isothermal phase of 60 s.
3. Irradiation of the specimen for 300 s at constant light intensity (5 mW/cm2 and 10 mW/cm2 ).
4. Isothermal phase of 60 s.
After finishing step four, the whole procedure is repeated with the cured specimen generating the baseline.
The baseline is subtracted from the heat flow signal of the first measurement to consider only the heat
flow of the crosslinking reaction. Otherwise, the light energy input would be taken into account.
In order to determine the maximum specific heat htot during the crosslinking reaction, an additional
measurement at non-isothermal temperature is performed. During 300 s irradiation, the temperature
is increased from 20 ◦ C to 70 ◦ C at 5 ◦ C/min leading to a maximum specific heat of htot = 319.73
J/g which is taken as reference value for the conversion of the specific heat during crosslinking reaction
into degree of cure for all measurements. The specific heat flows after substraction of the baseline of all
Table 1. Maximum attainable degree of cure qmax of all measurements

−10 ◦ C 0 ◦C 10 ◦ C 20 ◦ C 30 ◦ C 40 ◦ C 50 ◦ C
2
5 mW/cm 0.469 0.619 0.675 0.775 0.851 0.904 0.947
10 mW/cm2 0.593 0.685 0.775 0.845 0.907 0.949 0.997
measurements are shown in fig. 1. It can be clearly seen that the crosslinking reaction proceeds very
quickly and that the specific heat flows have reached their maximum value after only a few seconds.
Additionally, it is obvious that a higher light intensity leads to higher maximum values in the heat flows
and that a higher temperature accelerates the reaction. Moreover, complete curing is not achieved at low
temperatures and light intensities and the crosslinking reaction is stopped prematurely, cf. fig. 2. The
maximum attainable degrees of cure qmax for all measurements are listed in table 1. It can be seen that
an almost fully cured material is achieved at a temperature of 50 ◦ C, regardless of the light intensity.
This result must definitely be taken into account when formulating the material model for the degree of
cure.
Specific heat flow [mW/mg] Specific heat flow [mW/mg]
0
−10 0
−20 50 °C
−10 40 °C
−30 −20 30 °C
20 °C
−40 −30 10 °C
0 °C
−50 60 62 64 66 68 70 72 −10 °C
−60
0
−10 0
−10 50 °C
−20
−20 40 °C
−30 −30 30 °C
−40 20 °C
−40 10 °C
−50
0 °C
−50 60 62 64 66 68 70 72 −10 °C
−60
0 50 100 150 200 250 300 350 400 450
Time [s]
Figure 1. Measured specific heat flows after substraction of the baseline. Top: light intensity I = 5 mW/cm2 ,
bottom: light intensity I = 10 mW/cm2 .
176
Modelling of the degree of cure
The differential equation for the degree of cure q depends on time, light intensity I (z (t) , t) and temper-
ature T . Following the work of Kamal [3], an autocatalytic model of (m + n)th order incorporating the
maximum attainable degree of cure qmax is applied for the degree of cure q:
q̇ = (k1 (I (z (t) , t) , T ) + k2 (I (z (t) , t) , T ) q m ) (qmax − q) (2)

n
.
Here, I (z (t) , t) denotes the time- and location-dependent light intensity. The light intensity must depend
on the vertical location z (t) since it evolves during the printing process. The maximum attainable degree
of cure acts as a limit value. The constants
b
I (z (t) , t)
ki (I (z (t) , t) , T ) = k0i (T ) (3)
Iref
with the Arrhenius type equations

Ei
k0i (T ) = Ai exp − (4)
RT
are based on the proposal by Maffezzoli and Terzi [5] depending on light intensity I (z (t) , t) and temper-
ature T . If no light source is present, i.e. I (z (t) , t) = 0, the crosslinking reaction is stopped. Ai and Ei
are pre-exponential factors and activation energies, respectively. R denotes the universal gas constant.
The additional parameter b is introduced for a better representation of experimental data. Furthermore,
the auxiliary variable Iref = 1 mW/cm2 is used to convert the light intensity into a unitless value.
Parameter identification and validation
The parameters in eq. (2) are identified using the commerical optimisation program LS-OPT R
in com-
bination with MATLAB ’s ode45 solver. LS-OPT uses the sequential response surface method for
R R
optimisation. For further informations see [6].

The parameter identification depends strongly on the initial values. The parameters listed in table 2
represent the first result of the parameter identification which identified the equation for the degree of
cure on the measurement results. For simplification, the particular maximum attainable degree of cure
was specified as a parameter in order to achieve a better agreement with the measurement results. Fig.
2 shows the simulations of the degree of cure compared to all measurements according to eqs. (1) and
(2) with identified model parameters.
1.0
0.9
Degree of cure [−]
0.8
0.7 50 °C
0.6 40 °C
0.5 30 °C
0.4 20 °C
0.3 10 °C
0.2 0 °C
0.1 −10 °C
0.0
1.0
0.9
Degree of cure [−]
0.8
0.7 50 °C
0.6 40 °C
0.5 30 °C
0.4 20 °C
0.3 10 °C
0.2 0 °C
0.1 −10 °C
0.0
0 25 50 75 100 125 150 175 200 225 250 275 300 325 350
Time [s]
Figure 2. Measurement (—) and simulation (- -) of the degree of cure q with identified model parameters. Top:
light intensity I = 5 mW/cm2 , bottom: light intensity I = 10 mW/cm2 .
177
Table 2. Identified model parameters for the description of the crosslinking reaction
A1 A2 E1 E2 R m n b
1/s 1/s J/mol J/mol J/(mol K) - - -
0.6606 0.8744 9794 1001 8.314 4.554 2.996 1.365
Summary and outlook
First measurements for the experimental characterisation of the crosslinking reaction of photopolymers
used in additive manufacturing processes are presented in this contribution. It is shown that temperature
and light intensity have a significant influence on the crosslinking reaction.
A differential equation incorporating the maximum attainable degree of cure qmax is proposed for the
phenomenological description of the degree of cure. Parameter identification using commercial optimisa-
tion tools is carried out. The validation shows a good agreement between measurements and simulations.
Forthcoming considerations include the correct description of the maximum attainable degree of cure as
a function of temperature and light intensity:
qmax = f (T, I) (5)
Kolmeder et al., for example, have already carried out work for this purpose [4]. Since the glass transition
temperature of acrylic photopolymers strongly depends on the degree of cure, conventional DSC mea-
surements must be performed after irradiation to determine the glass transition temperature for distinct
degrees of cure. Subsequently, DiBenedetto’s proposal [1] can also be used to establish a relationship
between the maximum attainable degree of cure and the glass transition temperature.
Acknowledgement
The authors would like to thank the German Research Foundation (DFG) and Agence nationale de la
recherche (ANR) for the financial support of the project "Constitutive modeling of UV-curing printed
polymer composites".
References
[1] A. T. Di Benedetto. Prediction of the Glass Transition Temperature of Polymers: A Model Based
on the Principle of Corresponding States. Journal of Polymer Science: Part B: Polymer Physics 25
(1987), pp. 1949–1969.
[2] G. W. H. Höhne, W. F. Hemminger, and H.-J. Flammersheim. Differential Scanning Calorimetry.
Springer-Verlag Berlin Heidelberg, 2003.
[3] M. R. Kamal. Thermoset Characterization for Moldability Analysis. Polymer Engineering and Sci-
ence 14.3 (1974), pp. 231–239.
[4] S. Kolmeder, A. Lion, R. Landgraf, and J. Ihlemann. Thermophysical properties and material mod-
elling of acrylic bone cements used in vertebroplasty. Journal of Thermal Analysis and Calorimetry
105 (2011), pp. 705–718.
[5] A. Maffezzoli and R. Terzi. Effect of irraditation intensity on the isothermal photopolymerization
kinetics of acrylic resins for stereolithography. Thermochimica Acta 321 (1998), pp. 111–121.
[6] N. Stander, W. Roux, A. Basudhar, T. Eggleston, T. Goel, and K. Craig. LS-OPT
R
User’s Manual.
Livermore Software Technology Corporation. 2015.
178

Simulation of delamination in ferromagnetic-ferroelectric

laminate composites
1
Alexander Schlosser ∗
and Andreas Ricoeur1
1 University of Kassel
Abstract. A cohesive zone model for magnetoelectromechanical loadings is presented and the influence of
Maxwell stresses on the fracture behavior is investigated taking the example of a bimaterial DCB-specimen.
Furthermore, the influence of delamination in magnetoelectric laminates on their functional coupling properties
is shown.
1 Introduction
~
H ~tS
∆ϕel
~tmech
~tMaxwell
~ C, B
D ~C δtip
Crack Cohesive zone
Figure 1. Interface crack under magnetoelectromechanical loading and a cohesive model
The efficiency in converting magnetic into electric energy and vice versa makes magnetoelectric (ME)
composites promising candidates for many technical applications. The ferroelectric matrix as well as the
magnetostrictive inclusions of particle composites and the layers of laminates are mostly ceramics or other
brittle materials, thus being prone to cracking. Independent from the kind of composite the transmission
of stresses via the interfaces between the constituents plays the key role in the functionality of ME
composites. Therefore, the investigation of delamination processes is of great interest for the prediction
of durability and coupling factors. In order to investigate delamination processes in ME composites,
cohesive elements are being developed and applied in combination with ME finite elements described
in [1] and [2]. Those cohesive elements must be able to describe delamination fracture processes under
magnetoelectromechanical loadings as seen in Figure 1.
2.1 Magnetoelectroelastic behavior during delamination
The ME cohesive element is based on a potential
1 (∆ϕel )2 (∆ϕmag )2
Ψ = Ψmech − κ +µ − ei δi ∆ϕel − q i δi ∆ϕmag , (1)
2 δn + ∆ 0 δn + ∆ 0
∗ Corresponding author: alexander.schlosser@uni-kassel.de
179
where δi , ∆ϕel and ∆ϕmag are the discontinuities of the displacement, the electric potential and the
magnetic potential at the interface. δn , in particular, denotes the normal displacement jump. Here µ and
κ are the magnetic and electric permeabilities, respectively. The vectors q i and ei represent piezomagnetic
and -electric behaviors of the interface. The mechanical contribution Ψmech can be adapted to any
mechanical behavior of choice. In the following, a mechanical potential in the form of Ψmech = cij δi δj
is used, where cij is the cohesive stiffness tensor. The constitutive equations of the interface result from
the derivative of Eq. (1) with respect to the related independent variables. This leads to a composition
of five contributions to the restraining stress vector:
∂Ψ 1 ∆ϕel 2 ∆ϕmag 2
ti = = cij δj + κ +µ ni − q i ∆ϕmag − ei ∆ϕel . (2)
∂δi 2 δn + ∆ 0 δn + ∆ 0
The first contribution is the classical mechanical one resulting from the chosen traction-separation-law
(TSL). The second and third parts are the contributions of the electric and magnetic fields to the Maxwell-
stresses, followed by the linear piezomagnetic and -electric stress contributions. The derivations with
respect to the magnetic and electric potential jumps lead to the cohesive constitutive equations for the
magnetic surface normal Bi ni and the electric surface charge ωS :
∂Ψ ∆ϕmag
Bi ni = − = µ + q i δi
∂∆ϕmag δn + ∆ 0
and (3)
el
∂Ψ ∆ϕ
Di ni = −ωS = − =κ + ei δ i .
∂∆ϕel δn + ∆ 0
The first terms in the Eqs.(3) are an extended capacitor analogy, originally introduced in [3]. The
extension is on the one hand composed of a small constant ∆0 which is mainly incorporated to avoid
division by zero during numerical simulations. On the other hand, it is a dependency of the permeabilities
on the damage state of the interface (see below). The second terms, in turn, are the contributions of
piezomagnetic and - electric effects. All above mentioned material constants depend on a damage variable
f which is driven by the chosen TSL. This variable leads to a decrease of all material properties during
the fracture or delamination process, respectively. The properties cij , ei and q i decrease to zero when the
surface is completely damaged. Following [4], the permeabilities µ and κ decrease to the properties of a
chosen medium (e.g. air) filling the emerging crack.
3 Results
3.1 Fracture behavior under mode I/II/IV/V loading
A composite of one layer BaTiO3 (upper half) and one layer CoFe2 O4 (lower half) is investigated. Both
materials exhibit linear behavior and are perfectly poled in x2 -direction meaning that no switching pro-
cesses occur. Cohesive zone elements with the above described behavior are implemented between the
two layers. The piezoelectric and -magnetic properties of the cohesive zone are both neglected. The
permeabilities, including κBulk nad µBulk , are defined as averages between the upper and lower layer.
Two test cases have been simulated. In both cases first an electric field and a crack mouth opening dis-
placement uCOD are applied creating a Mode I/II/IV/V-loading. The Mode-II contribution arises from
the asymmetry of the specimen in combination with the piezoelectric effect. In the first set of simulations
the opening displacement was increased until the crack had grown through the whole specimen. In the
second case, a constant displacement is applied in a way that the crack only partly propagates through
the specimen. Figure 3 (Left) shows load-displacement-curves for different electric loadings ranging from
uCOD
2
BaTiO3 P~
∆ϕel
CoFe2 O4
uCOD
2 Cohesive zone elements
Figure 2. Two-layer-composite made of BaTiO3 and CoFe2 O4 and electromechanical loading
180
-1000 to 1000 V/mm. For positive signs, the electric field is parallel to the poling direction and for
negative ones it is antiparallel. One can clearly see a significant nonlinear influence of the electric field
V
0.6 E2∞ = ±1000 mm 7.8
V
E2∞ = ±500 mm
lcrack [mm]
V
0.4 E2∞ = ±200 mm 7.7
F [N]
V
E2∞ = 0 mm
0.2 7.6
tMax
i =0
0 7.5 tMax
i 6= 0
0 2 4 6 −1,000 −500 0 500 1,000
uCOD [mm] ·10−3 V
E2∞ /H2∞ [ mm A
/ mm ]
Figure 3. Left:load-displacement curves of the DCB-specimen with linearly increased opening displacements and
different electric fields. Right: crack lengths for a constant opening displacement uCOD = 1.0 · 10−3 mm
and different electric fields simulated with and without Maxwell stresses.
on the fracture behavior and at the same time an independece on the sign of the electric field. This leads
to the conclusion that, lacking piezoelectric properties of the cohesive or fracture process zone, the only
relevant influence of the electric load on the fracture behavior comes from the Maxwell stresses which are
proportional to the square of the electric field in the crack. The same conclusion can be drawn from the
results shown in the right plot of Figure 3. Here, the crack lengths for a constant opening displacement
uCOD = 1.0 · 10−3 mm and various electric fields simulated with and without Maxwell stresses are shown.
The crack lengths show a quadratic dependence on the electric field due to the Maxwell stresses. The
results without Maxwell stresses show a small ( 1%) dependency on the electric field which is caused
by the piezoelectric and -magnetic behaviors of the bulk materials.
3.2 Influence of interface cracks on ME coupling
As a second test specimen a three-layered composite in L-L configuration (both constituents are po-
larized/magnetized along the long edge of the laminate) was created. It also consists of CoFe2 O4 and
BaTiO3 with linear piezomagnetic and -electric behaviors. It is loaded with a magnetic field antiparallel
to the magnetization vector. Figure 4 shows the shear and normal stress distributions along a perfect
interface (no cohesive elements) under a large magnetic loading of H1 = −1000 A/mm. The stress distri-
bution is quite similar to one known from the free-edge effect (see [5]) and probably also mathematically
singular at the free edges. To investigate the influence of interface cracks on the ME coupling coefficient
g11 = HE1
1
, cohesive elements are placed along the interfaces. In order to mimic semipermeable cracks, the
damage variable f was locally set to f = 1. Figure 5 shows the results of these simulations for different
·10−2
H1
5
2]
CoFe2 O4 Mi hCFO
σij [ mm
N
H BaTiO3 Pi hBTO 0
x2
CoFe2 O4 Mi hCFO σ12
−5 σ22
x1
L 0 0.2 0.4 0.6 0.8 1
x1 /xmax [−]
hBTO
Figure 4. Left: three layered ME composite. H = 1 mm, L = 10 mm, H
= 0.5. Right: shear and normal
stresses along the interface
181
·10−9 H1 H
L1
1.5 Configuration I
Configuration II
Configuration III H1 H
L1
1 Configuration IV
H1 [ A ]
hE1 i V
Configuration V
Configuration VI
0.5 Configuration VII
H1 H
L1
0 H1
0 5 10 15 20
Overall crack length lcrack [mm]
L
Figure 5. Influence of crack lengths of different crack configurations on the ME coupling
crack configurations under a minor load of H1 = 1 A/mm to avoid crack growth. The coupling coefficient
decreases linearly with increasing crack lengths for the configurations I-IV and bilinearly for configuration
V. In case of inner cracks, the coupling coefficient is only significantly effected for lcrack ≈ L.
4 Conclusion
A cohesive zone element for delamination under magnetoelectromechanical loading has been developed
and implemented in the commercial FE software ABAQUS. The influence of Maxwell stresses on the
fracture behavior under mode I/II/IV/V loading has been shown at a bimaterial DCB-specimen. Fur-
thermore the influence of delamination on the MEcoupling of a three-layered laminate in L-L-configuration
has been investigated.
References
[1] Artjom Avakian and Andreas Ricoeur. Constitutive modeling of nonlinear reversible and irreversible
ferromagnetic behaviors and application to multiferroic composites. Journal of Intelligent Material
Systems and Structures 27.18 (2016), pp. 2536–2554.
[2] Artjom Avakian, Roman Gellmann, and Andreas Ricoeur. Nonlinear modeling and finite element
simulation of magnetoelectric coupling and residual stress in multiferroic composites. Acta Mechanica
226.8 (2015), pp. 2789–2806.
[3] Tian-Hu Hao and Zi-Yuan Shen. A new electric boundary condition of electric fracture mechanics
and its applications. Engineering Fracture Mechanics 47.6 (1994), pp. 793–802.
[4] Sergii Kozinov, Meinhard Kuna, and Stephan Roth. A cohesive zone model for the electromechan-
ical damage of piezoelectric/ferroelectric materials. Smart Materials and Structures 23.5 (2014),
p. 055024.
[5] Christian Mittelstedt and Wilfried Becker. Interlaminar stress concentrations in layered structures:
Part I-a selective literature survey on the free-edge effect since 1967. Journal of Composite Materials
38.12 (2004), pp. 1037–1062.
182

Simulation and Modeling of Stator Elastomers in Posi-

tive Displacement Motors
1
Alexander Schöne ∗
, Ralf Landgraf1 and Jörn Ihlemann1
1 Chair of Solid Mechanics, Chemnitz University of Technology, Chemnitz, Germany
Abstract.
Positive Displacement Motors contain an elastomer lining. During operation, the rubber is not only exposed
to mechanical loads, but also to volume changing effects. This loading leads to failure of the elastomer lining
and thus to failure of the whole motor. Consequently, it is necessary to consider the mentioned effects within
numerical simulations of the motor, which is possible by a corresponding extension of a material model for
the simulation of rubber like behavior. Obtained results show a significant influence of the volume changing
effects on the stresses and strains in the elastomer.
Introduction
Positive displacement motors (PDMs) are the most important drives for drilling operations in the oil and
gas industry. They deliver a high power density and are suitable for directional drilling [4]. Furthermore,
they are very economical and flexible in comparison to other types of downhole power units. The power
section of a PDM consists of a steel rotor and a contoured stator with an elastomer lining. Rotor and
stator have a different number of helical lobes, so that cavities between the two parts arise. Due to the
high pressure of the drilling fluid inside the drillstring, the cavities inside the power section move, which
leads to a rotation of the rotor inside the stator. Thus, the hydraulic energy of the fluid is converted into
mechanical energy of the rotating rotor, which drives the drill bit [1].
The purpose of the elastomer lining is to seal the cavities from each other and to ensure a tolerance re-
garding manufacturing inaccuracies. So it is of essential relevance for the whole power section. Moreover,
most failures of PDMs are due to failures of the stator elastomer. The rubber is not only exposed to the
mechanical loading due to the rotation of the rotor, but also to different volume changing effects. Beside
high temperature changes and huge hydrostatic pressures, also interactions with the drilling fluid and
additives contained therein cause significant volume changes. In order to model the resulting stresses and
strains inside the elastomer properly, it is necessary to take these effects into account.
1 Setup of the FE-model
In order to simulate the power section of a PDM with the use of the finite element method (FEM), an
appropriate model is required. This model should consider all important geometric and load information
and it should deliver highly resolved results in a reasonable amount of calculating time. One way to
achieve both of these contrary aspects is to simulate just a small part of the power section, which has
to be representative for the geometry and the load of the whole power section. Therefore, it is called a
representative model section (RMS).
The entire geometry would be received by a regular repetition of the RMS along the motor axis. To take
this repeatable structure of the sector into account, periodic boundary conditions (PBCs) are used. They
create couplings between the displacements of corresponding nodes on associated end faces by applying
∗ Corresponding author: alexander.schoene@mb.tu-chemnitz.de
183
coupling equations [3]. Thus it is guaranteed, that neither penetrations nor gaps occur between the
corresponding faces of adjacent RMSs.
Figure 1 depicts an RMS, which has been used for the simulation of the power section during operation.
The outer component in light gray is the contoured steel stator. On its inside is the elastomer lining of
even thickness (dark gray). The middle part is the steel rotor, which can rotate around the stator axis
and its own axis. It can be seen, that the rubber lining is meshed much finer than the other components.
The PBCs are applied on the front faces of all three parts and the corresponding faces at the back of the
RMS, which are not visible in the picture.
2 Material modeling
2.1 Modeling of inelastic rubber materials with the MORPH material model
For modeling the behavior of the polymer, the MORPH (Model of Rubber Phenomenology) material
model is used [2, 5]. It is a constitutive model to consider typical effects of filled rubber materials
undergoing large deformations, like nonlinearity, hysteresis and Mullins effects. The MORPH model
includes eight independent parameters p1 − p8 .
The equations for the MORPH are given in the box below. The total Cauchy stress σ is composed of
a Neo-Hookean like component, which depends on the material function α and the left Cauchy-Green
tensor b, an additional stress σ A , and a hydrostatic part including the hydrostatic pressure p and the
unit tensor I, where ()0 denotes the deviatoric part. The additional stress is constituted by an ordinary
∗
differential equation using the Zaremba-Jaumann derivative σ A . It depends on the material function β,
∗
the Tresca invariant of the Zaremba-Jaumann rate of the Cauchy-Green tensor b T and the limiting stress
σ L . The latter is calculated by an exponential tensor function and includes a further material function
γ. Moreover, the factor bST describes the maximum Tresca invariant of the Cauchy-Green tensor in the
time span [0, t]. It takes the pre-stretch into account and thus is a measure for the softening. To use the
MORPH for a specific material, the eight material parameters p1 −p8 have to be identified by appropriate
experiments.

 ∗

 ∗

A 0 ∗A ∗ b b T b
σ = 2αb + σ − pI σ = β b T σL − σ A
σ L = γ exp 
p7 ∗
S

 + p8 ∗
b T bT bT
bT = max [|bI − bJ | ; I, J = 1, 2, 3] bST = max [bT (τ ) ; 0 ≤ τ ≤ t]

 
p2 p4 p6
α = p1 + q 2 β=q 2 γ = p5 bST 1 − q 2

1+ p3 bST 1 + p3 bST p6 + 2 bST
For the usage of the MORPH material model in the context of an FE-simulation within the commercial
software ABAQUS, an appropriate implementation is required. Therefore, the available user interface
UMAT (User subroutine to define a material’s mechanical behavior) is used. The basis for the implemen-
tation of the MORPH material model is the strain rate tensor D. In comparison to an implementation
based on the deformation gradient F , the usage of D has the advantage of more stability and an extended
load range can be simulated.
2.2 Extension of MORPH with respect to volume changing effects
Because the implementation is based on the strain rate tensor, the extension for volume changing effects
uses the additive decomposition of the strain rate tensor. The modeled effects are chemical shrinkage
during the vulcanization process, thermal expansion due to temperature changes and swelling as a result
of the exposure of the stator elastomer to drilling fluids.
184
The modeling of volume changes is done by the introduction of a function ϕ, which represents the volume
ratio of corresponding deformations. The formulation of the deformation gradient due to volume changes
reads
F ϕ = ϕ1/3 I . (1)
With the help of the relation between deformation gradient and strain rate tensor in the form
. S
Dϕ = F ϕ · F −1
ϕ , (2)
the equivalent formulation for the strain rate tensor is given by
.
1ϕ
Dϕ = I. (3)
3ϕ
Taking into account the dependency of function ϕ from the temperature θ, the swelling variable q ∈ [0, 1]
and the shrinkage variable s ∈ [0, 1], equation (3) becomes:

11 ∂ϕ . ∂ϕ . ∂ϕ .
Dϕ = θ+ q+ s I. (4)
3ϕ ∂θ ∂q ∂s
To evaluate this equation, it is necessary to formulate a specific ansatz for the function ϕ, which considers
the three volume changing effects. Each of these aspects uses an exponential approach to calculate the
associated volume changes:
ϕθ (θ) = exp (αT (θ − θref )) ϕQ (q) = exp (αQ (q − qref )) ϕS (s) = exp (αS (s − sref )) . (5)
Within these equations, the different α-parameters represent constant volumetric expansion coefficients,
while θref , qref and sref are the reference values for temperature, swelling and shrinkage, respectively.
To consider all three effects in one simulation, different ways of connecting them and taking interactions
between them into account are thinkable. By now they have been treated as independent, but couplings
could easily be added.
On the basis of an implementation of the presented equations it is possible to simulate the volume
changing effects of rubber linings in PDMs.
3 Numerical results
The numerical simulations of the RMS shown in figure 1 have been carried out with the commercial
FE-software ABAQUS. At this point, a short overview about the general behavior of the elastomer inside
the power section will be provided based on the obtained simulation results.
The investigation of strains and stresses inside the rubber lining show repetitive distributions following
the movement of the rotating rotor in the stator. As expected, the hydrostatic pressure forms a maximum
line along the contact area between rotor and stator. Instead, the distribution of the von Mises stress
behaves significantly different, which is shown in figure 2. The picture displays a contour plot of the
stresses in the elastomer. Rotor and contoured stator, which have also been modeled, are not shown in
this picture.
It is clearly visible, that not one single maximum line occurs, but a combination of two maximum lines
separated by a minimum line. The reason for this behavior is, that rubber materials prefer shape changes
in comparison to volume changes. Consequently, the elastomer moves away from the direct contact zone
with the rotor. This explains the characteristic distribution of the von Mises stress, which takes the
deviatoric stresses into account. The minimum line in the middle of the two maximum lines is due to
nearly pure volumetric deformations in this area. So in total, with each rotation of the rotor, every
material point inside the elastomer undergoes a combination of a shape change, followed by a volume
change and a second shape change.
185
Figure 1. Meshed RMS for the simulation of the Figure 2. Distribution of von Mises stress in the elas-
drilling process tomer during operation
The described distribution occurs also if the volume changing effects are neglected in the simulation, but
the absolute values of stresses and strains are substantially smaller. So it is necessary to consider the
volumetric effect, to get realistic results for the mechanical load of the elastomer.
To understand the influence of different material properties, parameter variations have been carried out.
They show that the bulk modulus as well as the investigated volumetric expansion coefficients have a
strong impact on the resulting deformations, strains and stresses. So even little changes of those material
parameters lead to a significant change of the results, which demonstrates the importance of taking these
effects into account. Furthermore, the dependency of the results on the material parameters offers the
opportunity to influence the internal loads in the stator elastomer and thus its fatigue life.
Conclusions
The MORPH material model has been extended to take volume changing effects of the elastomer in PDMs
into account. The results of numerical simulations show very characteristic stress and strain distributions,
because rubber materials prefer shape changes compared to volume changes. In general, it is important to
consider the volume effects, because they have a significant influence on the mechanical load and thus on
the fatigue life of the stator elastomer. In future work, experiments should be performed to understand
the interactions between the described effects. Moreover, the vulcanization process could be regarded in
more detail, like given in [5]. In drilling mode, further aspects, like hysteresis heating and the resulting
thermal expansion, should be investigated. Finally, the model should be extended to take the external
loads in downhole operation into account, that have a big influence on the power section behavior.
References
[1] S. BA, M. Pushkarev, A. Kolyshkin, L. Song, L. L. Yin, et al. Positive Displacement Motor Modeling:
Skyrocketing the Way We Design, Select, and Operate Mud Motors. In: Abu Dhabi International
Petroleum Exhibition & Conference. Society of Petroleum Engineers. 2016.
[2] D. Besdo and J. Ihlemann. A phenomenological constitutive model for rubberlike materials and its
numerical applications. International Journal of plasticity 19.7 (2003), pp. 1019–1036.
[3] N. Goldberg, H. Donner, and J. Ihlemann. Evaluation of hyperelastic models for unidirectional short
fibre reinforced materials using a representative volume element with refined boundary conditions.
TECHNISCHE MECHANIK 35.2 (2015), pp. 80–99.
[4] T. Jaeger and J. L. Herlitzius. Navi-Drill Motor Handbook. Baker Hughes Incorporated, 2016.
[5] R. Landgraf and J. Ihlemann. Application and extension of the MORPH model to represent curing
phenomena in a PU based adhesive. In: Constitutive Models for Rubber X. CRC Press, 2017, pp. 137–
143.
186

Transient dissipation heating in polycrystalline ferro-

electrics accounting for mutual nonlinear couplings
1
Andreas Warkentin ∗
1 Institute
of Mechanics, Chair of Engineering Mechanics/ Continuum Mechanics, University of Kassel, Moenchebergstraße 7,
34125 Kassel, Germany
Abstract. This research is on irreversible domain wall motion, leading to a transient change in temperature
of the material, the so–called self–heating, if materials are subjected to high frequencies and large electrical
loads with insufficient heat dissipation. This work shows both the theoretical fundamentals and results of
the temperature change in ferroelectric materials assuming transient heating, as well as theoretical fundamen-
tals and results assuming an adiabatic process compared to experimental findings. Predictions concerning
polycrystalline material behavior are made by the condensed method.
Introduction
Ferroelectric materials, such as lead zirconate titanate (PZT) or barium titanate (BT) are technically
attractive ceramics because of their special properties. They are often used for actuators or sensors in the
precision range. Problems arise due to irreversible domain wall motion leading to self–heating associated
with changes in the material properties, thermal stresses and sometimes even phase transformations,
whereupon the devices finally are inoperative. In low Curie temperature materials, such as BT, even
depolarization is possible.
In order to be able to make predictions about the functionality in ferroelectric materials as a result of
temperature changes, it is necessary to be able to calculate the temperature change in ferroelectrics due
to irreversible domain–switching processes. For this purpose, the so–called condensed method (CM) is
extended by nonlinear bilateral caloric–electromechanical couplings, whereby the temperature change in
ferroelectric materials can be calculated based on a single polycrystalline material point representing the
entire material. The CM has been developed to simulate e.g. hysteresis loops and residual stresses for
polycrystalline materials without spatial discretization of the grain structure. It is suitable and efficient
for the implementation of different constitutive behaviors, considering interactions between grains or
different components of a material compound with low computational complexity and high numerical
stability [2].
In addition, a comparison of the results by this work and those of the FEM according to [4] is presented.
For further verification, the theoretical results from the CM and FEM are compared to experimental
results from [1].
1 Basic equations
In addition to the nonlinear constitutive equations of a thermoelectromechanical problem

σij = Cijkl εkl − εirr
kl − elij El − βij θ , (1)

Di = eikl εkl − εirr
kl + κij Ej + ki θ + Pi ,
irr
(2)

s = βij εij − εirr
ij + ki Ei + γθ , (3)
∗ Corresponding author: warkentin@uni-kassel.de ORCID iD: 0000-0001-5277-6327
187
where σij describes the mechanical stress, Di the electric displacement and s the specific entropy, and
the electromechanical balance laws σij,j = 0 and Di,i = 0, an energy balance is required to compute
the distribution of temperature change due to irreversible domain wall motions. The rate–dependent
formulation emanates from the local form of the first law of thermodynamics and the Gibb’s fundamental
equation of the chosen thermodynamical potential [4]:
A
−q̇j,j + ω̇ diss = T ṡ . (4)
Here, q̇jA is the surface heat flux and ω̇ diss is the dissipative power. Note, while T is the absolute
thermodynamic temperature of the material
θ = T − T0 (5)
denotes a temperature change. Inserting Eq. (3), in a rate dependent formulation, the energy balance
Eq. (4) results in:

A
−q̇j,j + ω̇ diss = T βij ε̇ij − ε̇irr
ij + ki Ėi + γ θ̇ . (6)
Summarizing the irreversibilities to χ̇ = ω̇ diss + βij ε̇irr

ij and neglecting reversible contributions to heating
or cooling in terms of the pyroelectric coefficient ki and the linear contribution of the thermal stress
coefficients βij yields
A
−q̇j,j + χ̇ = T γ θ̇ . (7)
The more general integral form of Eq. (7) is

Z Z Z
θ̇γT dV = χ̇ dV − q̇i ni dAS , (8)
Ω2 Ω1 ∂Ω
where Ω1 (region of the heat source) and Ω2 (volume around the heat source) are the considered areas
and ∂Ω is the surface of Ω2 .
1.1 Transient heating assuming a homogeneus temperature distribution and an adibatic behaviour
Inserting Fourier’s law of heat conduction into q̇j,j A

(note that the tensor of heat conduction λij is
assumed to be homogeneous in the body) in Eq. (7) and assuming a homogeneous distribution of temper-
ature change θ in a material point, the gradient θ,i and thus the derivative θ,ij disappear. This simplifies
the energy balance presuming an adiabatic behavior and it remains
θ̇γT = χ̇ . (9)
The solution of this differential equation in terms of to θ, using Eq. (5) resolved to T , is obtained as
follows:
Z θ1 Z χ1
∆χ
γ(T0 + θ)dθ = dχ = ∆χ ⇒ θ1 = + θ0 . (10)
θ0 χ0 γ T0 + 12 (θ1 + θ0 )
In Eq. (10), θ1 and θ0 are identified as temperature changes of the current and previous load steps and
∆χ as the associated dissipation. Inserting the leading order terms of the dissipative power ω̇ diss , further
assuming θ1 ≈ θ0 to obtain an explicit representation of the temperature from Eq. (10) and rewriting the
energy balance within the context of the CM [2, 3], the macroscopic multi–grain formulation of an i–th
time step is obtained as
1
θi = hσij ∆εirr irr irr
ij i + E i h∆Pi i + T hβij ∆εij i + θi−1 , (11)
T hγi
where angled bracketsPdenote quantities averaged over M grains of a representative volume element
M
(RVE), i.e. hξi := M
1
m=1 ξ , ∀m ∈ M and bars denote homogeneous quantities with the context of
m
a generalized Voigt approximation.
188
1.2 Transient heating with heat dissipation
In order to describe the transient heating, the heat flux q̇S = Θα with Θ = TS − T∞ has to be taken
into account. Inserting the heat flux relation into Eq. (8) by using the identity q̇iA ni = q̇S the differential
equation is obtained as
αAS VΩ1
θ̇ + θ = χ̇ , (12)
T {γi Vi } T {γi Vi }
Pv
where the abbreviation i=1 γi Vi =: {γi Vi } has been introduced with v as the number of all volumes
with dissmilar thermal properties. A further reduction with αAS /(T {γi Vi }) =: κ̃ and VΩ1 /(T {γi Vi }) =: β̃
provides
θ̇ + κ̃θ = β̃ χ̇ . (13)
The solution of the differential equation within the context of the CM [2, 3] provides the macroscopic
multi–grain formulation
VΩ1
θ(t) = hχ̇if g (t) = hχ̇i 1 − e−κ̃t , (14)
αAS
where f g (t) is a geometry– and time–dependent function, not depending on the calculations from the
CM.
2 Results
Fig. 1 shows the electromechanical loading schemes for the numerical examples which are shown in Fig. 2.
Starting from Eq. (11), Fig. 2 shows the results for the temperature development vs. load cycles compared
to the experimental results and those of the FEM.
experimental results [1]

FEM [4]
ext
2.5 Single crystal, σij =0
E2ext ext ext ext
σ22 , u2 CM, σij =0
2.0
Ê2pol CM, uext =0
θ [K]
1.5
Ê2
1.0
N
0 1 2 0.5
-Ê2 0.0
σ̂22 , û2
0 1 2 3 4
-Ê2pol
load cycles N [−]
Figure 1. Electromechanical loading schemes for Figure 2. Temperature development θ for poly-
the numerical examples crystal and single crystal polarized with
Ê2pol = 2 EC and loaded with Ê2 =
1.2 EC vs. electric load cycles N for dif-
ferent loading conditions.
The result from the simulation of a polycrystal, solid blue line, according to Eq. (11) is in very good
agreement with the experimental result [1] and those obtained by the FEM [4]. Due to the lack of residual
inter–grain stresses, the heating of a single crystal is larger than that of a polycrystal. On the other hand,
the temperature development is reduced by using displacement boundary conditions according to uext 2 = 0,
since it leads to increased external stresses. Based on the assumptions of the results shown in Fig. 2,
a stationary state is not attained, resulting in a monotonously increasing heating. Taking into account
temperature–dependent material parameters, the temperature change in the material approaches zero
with increasing heating and disappears when reaching the Curie temperature.
Calculated from Eq. (14), Fig. 4 shows the transient heating due to irreversible domain wall motions of
189
the sample shown in Fig. 3 induced by bipolar electrical load cycles with Ê2 = 1.2 EC . A stationary state
is attained depending on the geometry, the external electrical load and the frequency. Experiments are
work in progress and seem to confirm the shown results according to Fig. 4.
20.0
θ [K]
10.0
0.0
0 2,000 4,000 6,000 8,000
time t [s]
Figure 4. Temperature development θ for a poly-

Figure 3. Experimental setup
crystal. The material has been poled with
Ê2pol = 4.28 EC and cyclically loaded with
Ê2 = 1.2 EC . The load conditions ac-
cording to Fig. 1 with σ̂22 = û2 = 0 were
applyed
3 Conclusion
The results presented in this work show, that the extended CM is able to predict the temperature evolution
as a result of irreversible domain wall motion. The deviation between CM, experiment and FEM for a
larger number of load cycles is caused by the adiabatic boundary conditions. Taking into account heat
dissipation as well as temperature–dependent material parameters, a stationary state is finally attained.
References
[1] H.-S Chen and Y.-M Pei. Rate Dependent Heat Generation in Single Cycle of Domain Switching of
Lead Zirconate Titanate via In-Situ Spontaneous Temperture Measurment. Applied Physics Letters
102 (2013).
[2] A. Ricoeur and S. Lange. A Condensed Microelectromechanical Approach for Modeling Tetragonal
Ferroelectrics. International Journal of Solids and Structures 54 (2015), pp. 100–110.
[3] A. Ricoeur and S. Lange. Constitutive Modeling of Polycrystalline Multiconstituent and Multiphase
Ferroic Materials Based on a Condensed Approach. Archive of Applied Mechanics 89 (2019).
[4] Marius Wingen and Andreas Ricoeur. Caloric Aspects of Nonlinear Ferroelectric Constitutive Be-
havior: Modeling and Simulation. Continuum Mechanics and Thermodynamics 31 (2019), pp. 549–
568.
190
MS 09: Advances in full–field strain and temperature measurements for application in experimental mechanics

Numerical treatment of material parameter identifica-

tion using finite elements
Rose Rogin Gilbert ∗1
and Stefan Hartmann1
1 Institute of Applied Mechanics, Clausthal University of Technology, Germany
Abstract. In solid mechanics, different properties of materials are characterized by different constitutive
models, particularly models of evolutionary-type. The material parameters in these constitutive models are
calibrated to experimental data to predict the material behavior within the range of application and is called
material parameter identification, see [2, 5]. Different types of experiments are performed in order to uniquely
identify the material parameters. Reaction forces from the force gauges as well as surface strains or displace-
ments using Digital Image Correlation (DIC) methods are most widely used experimental data. The sensitiv-
ities of the material parameters are provided to the least-square solver to identify the material parameters.
There are different methods to obtain these material parameter sensitivites. External Numerical Differentiation
(END) is the most commonly used method to calculate these sensitivities, see, for the terminology [10]. How-
ever, for a complex constitutive model with inhomogeneous deformation having large number of parameters,
the computational time is high when using END. To circumvent this disadvantage, Internal Numerical Dif-
ferentiation (IND) is performed which reduces the computational time. Even though the computational time
is reduced, there are additional storage requirements that need to be fulfilled when performing IND method.
Introduction
There are many materials having different properties and behaviors, which are characterized in solid
mechanics by constitutive models. These constitutive models have material parameters, which have to be
calibrated to experimental data in order to predict the material behavior within the range of application.
This calibration is called material parameter identification or briefly parameter identification, see [2, 5].
Depending on the homogeneity of the deformation, the procedure of material parameter identification
varies. If the deformation is inhomogeneous, the entire boundary-value problem need to be solved, com-
monly by using finite element method. The material parameter identification for inelastic materials using
finite elements were discussed in [1, 7, 8]. To identify the material parameters, the sensitivities due to
parameter variations needs to be calculated. The most common method to calculate these sensitivites is
external numerical differentiation, see, for the terminology, [10]. However, for a constitutive model with
inhomogeneous deformation having large number of parameters, the computational cost is high. To cir-
cumvent this drawback, internal numerical differentiation is performed, which reduces the computational
time by a considerable amount.
Problem statement
The most commonly used basis for parameter identification process is a least-square formulation. The sim-
e
ulated results s ∈ Rns (nes is the number of simulated data) from the finite-element program are compared
e
with the experimental data d ∈ Rnd (ned is the number of experimental data from one or more exper-
iments). The residual vector, r̃(κ) = W {s(κ) − d}, describes the difference between the experimental
and the simulated data. In order to consider the different order of physical values (forces, displacements),
weighting matrices W are introduced. The so-called objective function, f (κ) = 21 r̃T (κ)r̃(κ), is then mini-
mized in a least-square sense using methods outlined in [3]. The displacements from the experiments are
obtained by using Digital Image Correlation (DIC), which are then projected onto the FEM surface data
to obtain the displacements at each evaluation points. The forces are obtained from force gauges of the
∗ Corresponding author: rose.rogin.gilbert@tu-clausthal.de
191
testing machnine. A trust-region reflective algorithm, provided by Matlab routine lsqnonlin.m, is used
to solve the problem. A necessary condition for a minimum at κ = κ∗ requires the derivative to be zero

df
∗
F(κ ) = = D T (κ∗ ) {s(κ∗ ) − d} = 0, (1)
dκ κ=κ∗
with D(κ) := dr̃(κ)/dκ = ds(κ)/dκ, D ∈ RnD ×nκ . D(κ) represent the sensitivites and are calculated
either by external differentiation or internal differentiation.
External Numerical Differentiation
It is more common to apply external differentiation schemes to approximate the sensitivites by
X n
D X n
dũn+1 κ
ũn+1 (κ + ∆κj ej ) − ũn+1 (κ)
≈ ei eTj , (2)
dκ i=1 j=1
∆κ j
with the “unit vectors” ei ∈ RnD (all entries are zero except for one having a 1 in row i) and ej ∈ Rnκ .
Here, the entire finite element program has to be run (nκ + 1)-times, for each iteration of the least-square
solver. For each iteration, the material parameters are perturbed in order to obtain the sensitivites using a
forward difference scheme. Therefore, the computational cost to calculate the sensitivities using external
numerical differentiation is very high when the number of parameters to be identified are high. If high
accuracy requirements are necessary, the external numerical differentiation might lead to an insufficient
iteration scheme to obtain a local minimum.
Internal Numerical Differentiation
In the case of internal numerical differentiation, the entire time integration step is assumed to be depen-
dent on the parameters κ. For pure hyperelatic constitutive models, the computation of the sensitivities
is simple and can be easily implemented, see for example [9],
∂g du ∂g du
g(t, u(κ), κ) = 0 =− Du = . (3)
∂u dκ ∂κ dκ
Once the equilibrium conditions, g = 0, are satisfied after each load step tn , the sensitivities Du |tn have
∂g
to be computed and stored. The quantity can be obtained by using the code generating software
∂κ
Acegen [6].
However, for constitutive models of rate-type (viscoelastic, viscoplastic) the internal numerical differ-
entiation is more complicated. The sensitivities are calculated based on Multilevel-Newton algorithm
and are provided from the finite element program itself. Following spatial discretization, a system of
differential-algebraic equations (DAE)
g(t, u(κ), q(κ), κ) = 0,

(4)
q̇(t, κ) = r(t, u(κ), q(κ), κ)
are obtained. Applying classical Backward-Euler method to the DAE-system leads to
g(tn+1 , un+1 (κ), qn+1 (κ), κ) = 0,

(5)
l(tn+1 , un+1 (κ), qn+1 (κ), qn (κ), κ) = 0,
with the integration step for the internal variables

l tn+1 , un+1 (κ), qn+1 (κ), qn (κ), κ = qn+1 (κ) − qn (κ) − ∆tn r tn+1 , un+1 (κ), qn+1 (κ), κ . (6)
For convenience, the following abbreviations
∂g ∂g ∂g ∂l ∂l ∂l
Gu = , Gq = , Gκ = , Lu = , Lq = , Lκ = . (7)
∂u ∂q ∂κ ∂u ∂q ∂κ
192
are introduced. The derivative of the (5) with respect to the parameters κ, which are the sensitivities,
at time tn+1 yields
Gun+1 Dun+1 + Gqn+1 Dqn+1 + Gκn+1 = 0
(8)
Lun+1 Dun+1 + Lqn+1 Dqn+1 + Lqn Dqn + Lκn+1 = 0.
Following some rearrangements, as shown in [4], (8)2 leads to
−1
Dqn+1 = −Lqn+1 [Lun+1 Dun+1 − Lqn Dqn + Lκn+1 ] . (9)
This is then inserted into (8)1 to obtain

h i
−1 −1
Gun+1 − Gqn+1 Lqn+1 Lun+1 Dun+1 = −Gκn+1 + Gqn+1 Lqn+1 − Dqn + Lκn+1 . (10)
The coefficient matrix on the left-hand side is exactly the tangential

stiffness matrix of the equilibrium
−1
iteration from the MLNA. The coefficient Lqn+1 − Dqn + Lκn+1 = Dqn+1 is the result of the linear
system, which has to be solved on Gauss-point level and has to be stored for the next time-step, in which
they represent the quantity Dqn for the next time-step. The total amount of storage is nel ×neGP ×nq ×nκ
where nel is the number of elements, neGP the number of Gauss-points in each element, and nq the number
of internal variables.
These calculated sensitivities are then provided to the Matlab routine lsqnonlin.m to identify the ma-
terial parameters. Since the sensivities are calculated internally using Acegen, the entire finite element
program has to be run only once per iteration of the least-square solver. This reduces the computational
cost considerably. However, there are additional storage requirements that needs to be fulfilled.
Conclusion
The basic concepts of material parameter identification along with different methods to obtain the sensi-
tivites are detailed in this short communication. Two methods, namely external numerical differentiation
and internal numerical differentiation, to obtain the sensitivities are explained. END is the most com-
mon method which is time consuming but easily implemented. IND on the other hand is a complicated
method but faster than END. The main drawback of the IND is the additional storage requirements. A
clear algorithmic structure is provided which can be extended according to different requirements.
References
[1] K. Andresen, S. Dannemeyer, H. Friebe, R. Mahnken, R. Ritter, and E. Stein. Parameteridentifika-
tion für ein plastisches Stoffgesetz mit FE-Methoden und Rasterverfahren. Bauingenieur 71 (1996),
pp. 21–31.
[2] J. V. Beck and K. J. Arnold. Parameter estimation in engineering science. 1st. New York: John
Wiley & Sons, 1977.
[3] J. E. Dennis and R. B. Schnabel. Numerical methods for unconstrained optimization and nonlinear
equations. Vol. 16. Classics in Applied Mathematics. Philadelphia: SIAM Society for Industrial and
Applied Mathematics, 1996.
[4] S. Hartmann. A remark on material parameter identification using finite elements based on con-
stitutive models of evolutionary-type. Computer Assisted Methods in Engineering and Science Vol.
24, no. 2 (2017), pp. 113–126.
[5] S. Hartmann and R. R. Gilbert. Identifiability of material parameters in solid mechanics. Archive
of Applied Mechanics 88.1 (2018), pp. 3–26.
[6] J. Korelc. Multi-language and multi-environment generation of nonlinear finite element codes. En-
gineering with Computers 18.4 (2002), pp. 312–327.
[7] R. Mahnken and E. Stein. A unified approach for parameter identification of inelastic material
models in the frame of the finite element method. Computer Methods in Applied Mechanics and
Engineering 136 (1996), pp. 225–258.
[8] R. Mahnken and E. Stein. Parameter identification for finite deformation elasto-plasticity in prin-
cipal directions. Computer Methods in Applied Mechanics and Engineering 147 (1997), pp. 17–39.
193
[9] G. Scheday. “Theorie und Numerik der Parameteridentifikation von Materialmodellen der finiten
Elastizität und Inelastizität auf der Grundlage optischer Feldmessmethoden”. PhD-thesis, Report
No. I-11 (2003). Institute of Mechanics: University of Stuttgart (Germany), 2003.
[10] K. Schittkowski. Numerical data fitting in dynamical systems. Dordrecht: Kluwer Academic Publ.,
2002.
194

Parameter identication for constitutive models of tex-

tile composite materials using digital image correlation
∗1 2 1 2 1
Justin Felix Hofmann , Claudia von Boyneburgk , Sophie Tunger , Hans-Peter Heim and Detlef Kuhl
1 Institute of Mechanics and Dynamics, University of Kassel

2 Institute of Material Engineering, Polymer Engineering, University of Kassel
Abstract. This contribution focuses on a composite material consisting of multiple woven layers made of
continuous wooden strips (salix americana) embedded in a polymer matrix (polypropylene). To asses the
suitability of the material in an architectural context, a nite element simulation of potential use-cases is carried
out. The simulation is based on a constitutive model assuming homogenous orthotropic elastic behaviour and
will be expanded to include multiscale modelling as well as nonlinearity in future research.
The non-isotropic behaviour as well as the heterogeneous structure of the material lead to a signicant inhomo-
geneity of the resulting stress and strain distributions during experiments, especially in o-axis congurations.
This can't be suciently captured using traditional measurement techniques. However, the advances in full-
eld measurement techniques enable us to capture this behaviour in suciently large regions of the specimen.
Experimental results of uniaxial on- and o-axis tensile tests using digital image correlation are shown and
used for the identication of the material parameters of the given model. Two dierent identication processes
and their transferability to more complex material models are discussed.
Motivation
Standardized testing methods for the determination of elastic material parameters often focus on isotropic
mechanical behavior. They require specic specimen geometries to generate homogeneous strain states
during testing, which in turn require only a small number of local measurements to be fully captured.
However, nonisotropic, as well as inhomogeneous materials demand more sophisticated measurement
techniques, which in turn require more complex processing. The parameter identication methods pre-
sented here are based on nite element model updating and process full-eld displacement measurements
gained during experiments. Rectangular specimens (Fig. 2) were tested in on and o-axis congurations
in a quasi static uniaxial tensile test (Fig. 1). However, the methods could be applied to other specimen
geometries and forms of testing as well.
The methods are applied to a composite material based on woven fabrics made from continuous wooden
bers and thin polypropylene lms. It is expected that the material behavior is orthotropic.
1 Experiment
To enable independent stress-strain states during the tests, the specimens were cut from the base material
at dierent angles (0◦ /45◦ /90◦ ) with respect to the warp direction of the textile layers. Pictures of the
specimens surfaces were taken at regular intervals during the experiments using a two-camera setup
and converted into displacement data uexp using digital image correlation. The transformation between
camera and specimen coordinate systems was acquired during a calibration step. Corresponding force
measurements were taken using a load cell.
The experimental results have demonstrated highly inhomogeneous displacement elds (Fig. 4) due to
the inhomogeneity of the composite.
∗ Corresponding author: hofmann@uni-kassel.de
195
Figure 2. 45◦ -specimen
Figure 1. Uniaxial tensile tests Figure 3. Finite element mesh
u1 [mm]
x2 [mm]
x2 [mm]
ε11 [−]
x1 [mm] x1 [mm]
Figure 4. Displacement eld (45 -specimen)

◦
Figure 5. Strain eld (45◦ -specimen)
2 Finite element model
A static nite element model reproducing the test setup was created separately for each specimen and load
step. Each model includes the specimen geometry (width, thickness, eective length between clamps),
the orientation of the material with regard to the specimen coordinate system, as well as the boundary
conditions applied during the campaign (clamps, force). The specimens were discretized using a nite
element mesh consisting of bilinear quadrilateral elements (Fig. 3). The initial material parameters p0
for an orthotropic linear elastic material model were chosen arbitrarily.
3 Parameter identication
Two parameter identication algorithms were implemented and the results were compared. The rst
algorithm is based on [1], the second is a modied version of this method. Both follow the idea of nite
element model updating (Fig. 6 and Fig 7). Experimental and numerical results are compared on the
strain level using error function
PNi
M
X num
j=Ni−1 +1 |εj (p) − εexp
j |
2
e(p) = PNi exp 2
i=1 j=Ni−1 +1 |εj |
where εnum
j and εexp
j are pairs of spatially corresponding strains, which are grouped into M subsets (one
for each specimen and load step). The error is subsequently minimized using the gradient descent method
to nd a set of optimal material parameters. The dierence of both versions is in the computation of the
strain eld εexp from uexp , which occurs prior to solving the inverse optimization problem.
The rst algorithm utilizes the nite element mesh of the corresponding nite element simulation to ap-
proximate experimental strain data at the gauss points of each element containing sucient corresponding
measurements by solving a local element-wise optimization problem described in [1]. The quality of the
strain computation is therefor dependent on the chosen nite element mesh and spatial discretization of
the nite element model is limited by the resolution of the measured displacement eld.
The second algorithm generates strain data independently of the nite element simulation. Delaunay
triangulation is applied to each set of measuring points, to generate a mesh consisting of triangular
196
Figure 6. Algorithm of [1] Figure 7. Modied algorithm
old
6 old 0.6 6 old 6
modied 0.6
modied modied
E1 E1
[GP a]
4
4
[−]
[GP a]
E2 0.4 E2 0.4
[−]
[−]
2 ν12
2 ν12
2
0.2
0.2
G12
G12
0
0 2 4 6 8 10
0 2 4 6 8 10 0 2 4 6 8 10
element size [mm]
number of iterations number of iterations
Figure 10. Converged solutions for

Figure 8. Model parameters Figure 9. Error function
various element sizes
elements with known node displacements uexp . Strain data εexp is subsequently calculated at the gauss
point of each element (Fig. 5).
Both algorithms were tested for dierent starting values and element sizes. Given reasonable starting
values, both algorithms fully converged after approximately 10 iterations (Fig. 8 and 9). The converged
solutions were independent of the starting values. While the rst algorithm required much less com-
putation time, it showed a much higher dependency on the chosen nite element size, compared to the
modied version (Fig. 10).
4 Conclusion and outlook
Two versions of an inverse method for the identication of the elastic material parameters of orthotropic
materials have been presented. The methods are based on nite element model updating. Displacement
measurements taken during tensile tests were compared to the results of a nite element simulation using
a strain based error function, which is subsequently minimized via the gradient descent method. It could
be shown that the methods converge against solutions for the orthotropic material parameters, which
are in agreement with the expected range of values, based on previously determined material parameters
of the individual material components and the theory of mixtures of composites. The solutions are
independent of the chosen starting values. However, the rst algorithm showed a high dependency on
spatial discretization.
The experimental results show a strong material inhomogeneity, which can't be reproduced by the material
model used in this work. Consequently, the parameter identication method will be applied to multiscale
modeling in future research.
References
[1] D. Lecompte, H. Sol, J.Vantomme, and A. M. Habraken. Identication of Elastic Orthotropic Ma-
terial Parameters Based on ESPI Measurements. SEM annual conference and exposition on experi-
mentaland applied mechanics (2005).
197
198

Monitoring of multiaxial failure surfaces and the evo-

lution of deformation bands in PVC foams using 3D
digital image correlation
Anne Jung
∗ 1, Francesca Concas
1 and Stefan Diebels
1
1 Chair of Applied Mechanics, Saarland University, Saarbrücken, Germany
Abstract. Cellular materials such as foams are biomimetic structures used for lightweight constructions and
for energy absorption. Closed-cell polyvinylchloride (PVC) foams are among the most used core materials for
sandwich composites. Therefore, a deep knowledge of their multiaxial behaviour is fundamental for application
of this material. The present contribution focuses on the evolution of failure surfaces and the elucidation of the
local deformation mechanism under multiaxial loading. Uniaxial,pure torsion as well as torsion experiments
with superimposed uniaxial loads combined with 3D digital image correlation are performed.
1 Introduction
Cellular materials are built according to the blueprint of nature as lightweight construction materials with
porous structure. Open-cell foams are inspired by bones, while closed-cell foams mimic the structure of
wood or cork [1]. Although metal foams oer advantages concerning ductility and energy absorption
capacity, they lack in terms of corrosion and elasticity. In contrast, polymer foams can be used advan-
tageously due to their lightweight construction properties and exibility. However, they lack in terms of
stiness and thermal stability. Usually, foams are processed in the form of sandwich structures with foam
core. Closed-cell polyvinylchloride (PVC) foams are among the most used core materials for sandwich
composites. Applications are in ship building as well as in aerospace industry and for wind turbine blades
[1]. In all of these applications, multiaxial loading is of utmost importance. Hence, the investigation of
failure surfaces under multiaxial loading is needed for the secure design of components made of closed-
cell PVC foams. However, up to now, experimental investigations on cellular materials under multiaxial
loading are very limited due to the challenging and complex nature of the experiments.
The present contribution deals with the experimental multiaxial characterisation of a closed-cell PVC
foam and the probing of the corresponding failure surfaces. The local deformation mechanisms were
studied by 3D digital image correlation (DIC) using an eight camera system.
2 Material and Methods
Closed-cell PVC foams were purchased from Diab International (Laholm, Sweden). Especially plates of
Divinycellr HP100 with a density of 100 g/cm3 were studied. Cylindrical dog-bone specimens with
a height to diameter ratio of 2:1 were cut from in-plane direction of the plates. More details on the
specimen preparation can be found in Concas et al. [3].
Multiaxial experiments under dierent loading conditions are needed for probing of the failure sur-
face. Therefore, uniaxial compression and tension experiments, torsion experiments and torsion ex-
periments with superimposed constant uniaxial tensile or compression loads were performed using an
Instron ElectroPulsr E10000 universal testing machine (Instron Ltd., Pfungstadt, Germany). A 360◦
DIC analysis was carried out to elucidate the local deformation mechanism under multiaxial loading.
Therefore, an eight camera system and the commercial DIC software Istra4Dr (Dantec Dynamics Ltd.,
Slovkunde, Denmark) were used. A detailed description of the procedure for the experimental probing
of failure surfaces for foams can be found in previous works [3, 5, 6].
∗ Corresponding author: anne.jung@mx.uni-saarland.de
199
3 Results and Discussion
Figure 1 presents the experimentally probed failure surface for the Divinycellr HP100 foams in the
hydrostatic plane, which is formed in the stress space by the hydrostatic axis and the deviatoric axis.
The hydrostatic axis is dened by the rst principal invariant I1 of the stress tensor. The deviatoric axis
describes pure shear states and is dened by the square root of the negative second deviatoric invariant
J2 of the stress tensor. The resulting asymmetric closed failure surface is shifted along the hydrostatic
axis to the tensile part and has its maximum under combined tension-torsion loading. Pure tensile failure
stresses are higher than pure compression failure stresses. The closed shape of the failure surface accounts
for the compressibility of the foam.
A single-surface yield criterion to reliably describe such asymmetric failure surfaces is the yield criterion
according to Bier et al. [2], which was introduced for multiaxial yielding of compacted powder. However,
the t in Figure 1 shows the good approximation of the failure surface by this criterion for the inves-
tigated Divinycellr HP100 foams. The corresponding material parameters can be identied by inverse
calculations [4].
r
Figure 1. Failure surface of a closed-cell Divinycell HP100 foam.
The DIC analysis of the individual experiments outlined buckling during the plateau phase under com-
pression loading making a further investigation of the mechanical properties dicult. Uniaxial tensile
loading results in a concentrated failure by a crack perpendicular to the loading direction. Failure under
torsion results in crack with a 45◦ angle to the loading direction. The torsion experiments with superim-
posed uniaxial loading result in localisation bands in two diametrically opposite regions of the specimen,
where the highest strain gradients occur at the beginning of fracture. The corresponding DIC analysis
can be found by Concas et al. [3].
4 Conclusion
The present contribution dealt with the experimental probing of a failure surface for closed-cell PVC
foams and the monitoring of the local deformation using 3D DIC. An asymmetric, closed failure surface
shifted to higher tensile failure stresses was received. It outlined the importance of multiaxial experiments
for the reliable design of components made of such foams. The 360◦ DIC provided important and
interesting insights in the local deformation mechanisms and the evolution of strain elds during uniaxial
and multiaxial loading.
200
References
[1] M. F. Ashby. The mechanical properties of cellular solids. Metallurgical Transactions A 14.9 (1983),
pp. 17551769.
[2] W. Bier and S. Hartmann. A nite strain constitutive model for metal powder compaction using
a unique and convex single surface yield function. European Journal of Mechanics-A/Solids 25.6
(2006), pp. 10091030.
[3] F. Concas, S. Diebels, and A. Jung. Multiaxial failure surface of PVC foams and monitoring of
deformation bands by three-dimensional digital image correlation. Journal of the Mechanics and
Physics of Solids 130 (2019), pp. 195215.
[4] S Heinze, T Bleistein, A Düster, S Diebels, and A Jung. Experimental and numerical investigation of
single pores for identication of eective metal foams properties. ZAMM-Journal of Applied Mathe-
matics and Mechanics/Zeitschrift für Angewandte Mathematik und Mechanik 98.5 (2018), pp. 682
695.
[5] A Jung and S Diebels. Microstructural characterisation and experimental determination of a multi-
axial yield surface for open-cell aluminium foams. Materials & Design 131 (2017), pp. 252264.
[6] A Jung and S Diebels. Yield surfaces for solid foams: A review on experimental characterization and
modeling. GAMM-Mitteilungen 41.2 (2018), e201800002.
201
202

Full-field strain measurements on single struts of a Ni/PU

hybrid foam
Martin Reis∗1 , Stefan Diebels1 and Anne Jung1
1 Chair of Applied Mechanics, Saarland University, Saarbrücken, Germany
Abstract. Ni/PU hybrid foams are an example of open-cell hybrid foams. They are based on a polyurethane
(PU) foam template coated with nanocrystalline nickel (Ni). Due to their coating they show quite impressive
energy absorption capacity at low weights. Due to this fact, they have a high potential as energy absorbers.
The deformation in the macroscale results from single strut deformation in the microscale. Therefore, the
study of each size scale is necessary to get a full understanding of the material. In the present contribution,
tensile tests on individual struts of a Ni/PU hybrid foam are presented and evaluated using a homogenised
geometrical model of each strut. To proof the reliability of the homogenised models and to analyse the coating
thickness, micro sections of five individual struts are carried out. The micro sections show a clear difference
in coating thickness but almost no scattering of the cross sectional area for the PU core. This strengthened
the use of a homogenised geometrical model.
1 Introduction
Foams are bio-inspired structures, which can be divided into closed-cell foams, comparable to the inner
structure of cork, and open-cell foams, inspired by trabecular bones. In the present work, open-cell foams
are of interest. Thereby, the macroscopic structure consists of single pores on the mesoscale and single
struts on the microscale. Through the open structure, these materials allow to be coated by an electrode-
position process with different kinds of metals. Therefore, an open porous template, such as aluminium
or the cheaper PU, is needed. In the case of PU, an additional electric conductive coating is necessary to
enable the electrodeposition process. Coated foams are classified as hybrid foams. The material behaviour
of hybrid foams is dominated by the material behaviour of the coating material. Coating a PU foam
with Ni immensely increases the energy absorption capacity of the structure. Comparing a Ni/PU hybrid
foam with a pure aluminium foam shows that the energy absorption to weight ratio strongly increases
and the production costs decrease. Among others, both characteristics determine the material to be an
ideal lightweight structure with high potential for the use in the automotive sector and for aerospace
applications. Before it can be used in a real-world application, further investigations of the material
behaviour are necessary. Due to the micro structure, the material behaviour of the macroscopic foam is
composed of the behaviour of the pores and struts. Thus, realistic models require information from all
scales. In addition, the fine grain structure and different orientations of the grains in the coating affect
the material parameters of the Ni coating which differ from the material parameters of the bulk material.
Therefore, experimental investigations on single pores and struts are needed. The small dimensions of
the struts with a length of about 2 mm and a width of about 0.4 mm combined with the irregular surface
only allow using digital image correlation (DIC) to evaluate strains on the specimen’s surface. Due to
the inhomogeneous coating thickness and the variable strut geometry, the experimental results include
large scattering. A realistic stiffness can be determined by transferring the force-displacement-data to
a stress-strain-curve using the cross sectional area in the crack area and the initial length of the struts.
To avoid any additional uncertainties produced by the experimental set-up, high measurement accuracy
is important. Therefore, the DIC noise is minimised with a pre-test routine and the noise level is de-
termined for each experiment. This contribution presents results from tensile tests on single struts and
optical analysis of micro sections on single struts.
∗ Corresponding author: martin.reis@uni-saarland.de
203
2 Material and Methods
The following microscale experiments are performed with a custom-build testing device [jung2015]. The
used stepper motor in combination with the high precision screw drive allows a stroke resolution of 1
µm and a force up to 6 kN with a repeatability of 0.02 µm, controlled with a custom software based on
LabVIEWTM . Besides the controlling, the software records the forces and captures the images for the
DIC measurements. Due to the complex and individual structure of each strut, the specific geometry
of each strut is needed to calculated the stress. For this task, another custom-build device is used.
The set-up rotates a single strut in 24 steps ones around its longitudinal axis and acquires an image in
each step. This process is repeated for a second translation angle of the camera. Below the sample a
pattern allows to identify the rotation angle of the specimen. The information on the rotation of the
pattern combined with the edges and shadows of the sample found in each image allow to generate a 3D
geometric model using the commercial software 3DSOMTM . The resolution of this process is comparable
with the accuracy of a model received from computed tomography [bleistein2019]. Starting with this
model, the initial cross sectional area at the crack position can be measured. Therefore, the model is
distinguished in a stack of approximately 1,000 images. In this stack, the position of the crack is found
and the cross sectional area is measured for 11 images close to the crack position. Here the mean area is
calculated as fracture area. The initial area and the force allow to calculate the engineering stress during
the experiment. Even using a commercial software to perform a DIC measurement, many factors, as for
examples lighting, environment and subset size, have an influence on measurement noise. To achieve a
high quality measurement it is necessary to estimate the error and to reduce it to a minimum. For this
purpose, a simple three-step routine was evolved [reis2019]. In the first step, a series of images from
a static specimen is taken. For this series the DIC measured displacement, which equals the noise of
this experiment, is evaluated. The value is reduced to a minimum changing lighting settings, reducing
environmental influences and finding the optimal evaluation settings in the DIC software. These hardware
and software settings are validated in the second step with the measurement of a rigid body motion and
the comparison to the machine stroke. In the third step, a calibration target found by Patterson et al.
[patterson2007], adapted to the small dimensions of the present sample, was used. Due to its three-
point bending structure, it links the force measured by a force to the DIC measured displacement using
the Euler-Bernoulli beam theory. Both validation steps show, that the best settings in step 1 are the
best settings for the following measurements. 20 images are taken without any movement to quantify
the DIC noise for each test. A 9 MPx CCD camera (Manta G917B, Allied Vision Technology GmbH,
Stadtroda, Germany) in combination with a telecentric lens (DTCM16.6-110, vico imaging, Guangdong,
China) is used for imaging. The camera lens combination has a field of view of 12.8 × 10 mm2 with
a pixel edge length of 3.69 µm. Because of the coating process, the coating thickness locally varies in
the macroscopic foam sample. Since the individual struts are extracted from all over the foam sample,
their coating thickness varies as well. To measure the specific coating thickness, micro sections of five
exemplary struts are done. Therefore, the struts are moulded upright in a slip-free casting compound
(Technovit 4002IQ, Kulzer GmbH, Hanau, Germany) and polished with sandpaper (P1200, P2500) to
generate a micro section approximated every 50 µm in axial direction of the strut reaching up to 40
images. Due to processing reasons, only 33 micro sections could be made. The images were captured
with a light microscope (Axio Imager 2, Carl Zeiss Microscopy GmbH, Jena, Germany) using a lens with
a magnification of ×10 (EC "Epiplan-Neofluar" 10x/0,25 HD DIC M27, Carl Zeiss Microscopy GmbH,
Jena, Germany). For all struts in all images, the coating thickness was measured at six points (Figure 2
(a)). Additionally, the area of the inner PU core (Figure 2 (b)) was estimated using Heron’s formula (1)
to calculate the surface area of a triangle.
p
A = s(s − a)(s − b)(s − c) (1)
a+b+c
with s = and a, b, c : triangle side lengths
2
204
3 Results and Discussion
The deformation of the strut in the tensile test starts with a homogeneous strain distribution over the
complete strut during the linear elastic deformation (Figure 1 1 and ).
2 With damage initialisation,
first deformation bands occur (Figure 1 3 and )
4 and grow until the sample fractures (Figure1 5
6 Thereby, the orientations of the deformation bands show every case between pure tension and
and ).
shear deformation.
Figure 1. Full-field strain measurement using DIC on an individual strut during tensile testing at six different
1
deformation steps (left) and stress-strain-diagram with highlighted deformation steps - 6 (right)
The varying deformation behaviour could have different reasons. Due to the complex structure of the
struts, a perfect alignment in the device is not always possible. The error in the alignment changes the
experimental mode from pure tensile loading to bending the strut followed by a tensile load. Thereby,
the bending results in a local elevation of the load followed by an inhomogeneous failure. A second
reason can be found within the coating of the struts. The coating thickness varies around the strut,
as well as it contains defects which reduce the stability. Defects in the coating could only be located
using an in-situ CT scan, which is not available for the investigated struts. Therefore, a huge number
of experiments is needed to estimate homogeneous material parameters to map the real behaviour of the
individual struts. The scattering in stiffness for eight struts (TS1-TS8) can be found in Table 1. No
clear distribution function can be extracted, more experiments are needed to analyse the scattering of
the material parameters and to estimate a distribution function of the material parameters.
Table 1. Stiffness during linear elastic deformation for eight struts

TS1 TS2 TS3 TS4 TS5 TS6 TS7 TS8
stiffness [GPa] 31.766 27.873 52.672 40.766 42.184 48.385 68.414 24.249
For performing the micro sections, five struts (UD1-UD5) are chosen. During the investigation of the
coating thickness, a huge scattering can be found. For the five struts, the coating thickness scatters
from 20.96 ± 2.66 µm to 80.97 ± 10.22 µm. Whereas the cross sectional area of the PU core area stays
constant for all struts in all measurements (Figure 2). Hence, the outer dimensions of the strut in the
photogrammetric model allow to estimate the coating thickness.
4 Conclusion
Regarding the micro sections of struts it has been shown that the PU template foam has constant strut
cross sectional area for the different struts. Therefore, a good estimation of the coating thickness can
205
206
MS 10: Dynamic Systems with Non–standard Constraints: Simulation, Control and Numerical Analysis

A Lie group generalization of the RATTLE scheme ap-

plied to nonholonomic constraints
∗
Stefan Hante
Institute of Mathematics
Martin Luther University Halle-Wittenberg
06099 Halle (Saale), Germany
Abstract. We present an integration scheme which generalizes the well-known RATTLE scheme in two respects:
First, we will use a nonlinear configuration space with Lie group structure using a Lie algebra approach.
Second, we will extend the scheme to nonholonomic constraints in Pfaffian form. By using the approach
of variational integrators, we obtain a scheme which simultaneously supports holonomic and nonholonomic
constraints. We performed numerical experiments on the rolling disk problem, where the disk is considered a
rigid body with configuration space similar to SE(3). We will show test results and conclude that nonholonomic
constraints such as rolling or sliding contact can be incorporated in multibody simulation environments, where
it is infeasible to choose the best suited coordinates for each body by hand.
Introduction
When simulating large multi-body systems, it is impractical to choose minimal coordinates for each body
by hand. Rather, it is convenient to use standardized coordinates for position and orientation for all
bodies and introduce joints and restrictions by constraint equations. It is well-known that representing
orientation by a set of three rotation angles can lead to singularities if large rotations are present. Thus,
one has to use a non-linear configuration space for orientation. Fortunately, such non-linear configura-
tion spaces often possess Lie group structure. This can be used to construct integration schemes that
automatically preserve the structure of the non-linear configuration space. Here, we use the Lie algebra
approach [2] to construct a variational integrator [7], which works on a Lie group. The nonholonomic
RATTLie scheme is derived in order to solve equations of motion of mechanical systems with simultane-
ous holonomic and nonholonomic constraints. A nonholonomic RATTLE scheme was already introduced
in [4], which inspired this nonholonomic RATTLie scheme.
1 Lie group structured configuration spaces, variational principles and nonholonomic contraints
We consider a system with an n-dimensional configuration space G. We suppose G has a (matrix) Lie
group structure (G, ◦), where ◦ is the Lie group operation ◦ : G × G → G. Let e ∈ G be the identity
element. We call the tangent space Te G the Lie algebra of G. Since tangent spaces are linear, we can
identify Te G with Rn by an isomorphism e • : Rn → Te G. We define the hat operator b
• : Rn → Rn×n by
bg
a a, e
· b = [e b], where [•, •] is the Lie bracket of the Lie algebra.
The main idea of the Lie algebra approach [2] is to describe all derivatives q̇(t) ∈ Tq(t) G for a function
q(t) ∈ G by velocity vectors v(t) ∈ Rn and the kinematic equation
g
q̇(t) = dLq(t) (e) v(t), (1)
∗ Corresponding author: stefan.hante@mathematik.uni-halle.de
207
where dLq is the differential of the left multiplication Lq : Q 7→ q ◦ Q. The exponential function
exp : Te G → G is defined by the evolution of (1) with constant velocity vector w:
e = q(t)
exp(tw) ⇔ e
q̇(t) = dLq(t) (e) w, q(0) = e.
The tangent operator T : Rn → Rn×n characterizes the derivative of the exponential function:
d
exp(e ^v̇,
v ) = dLexp(ev) (e) T(v)· v = v(t), v̇ = v̇(t) .
dt
The derivative with respect to an element of the Lie group G can be expressed in our framework using
the differential operator D defined by

e
Df (g) · w = df (g) dLg (e) w, g ∈ G, w ∈ Rn , Df (g) ∈ Rdim(im f )×n .
For Lie groups which typically appear in multi-body dynamics (such as products and semidirect products
of R3 , unit quaternions S3 , etc.) these Lie group functions exp, T, e
•, b
•, etc. can be expressed in closed
form and are easy to compute.
We will consider a constrained mechanical system with Lie group structured configuration space G 3 q(t),
T
kinetic energy T v(t) = 21 v(t) · M · v(t) and potential energy U q(t) , where q(t) and v(t) are related
through the kinematic equation (1) and M is a constant, symmetric and positive definite
mass matrix.
We assume the mathematical system is constrained by holonomic constraints Φ q(t) = 0 as well as by
nonholonomic constraints in Pfaffian form B q(t) · v(t) = 0 with Φ(q) ∈ R m1
and B(q) ∈ Rm2 ×n . We
use the Lagrange-d’Alembert principle with an augmented Lagrangian for the holonomic constraints and
arrive at the constrained variational principle
Z T
T
0=δ T v(t) − U q(t) − λ(t) · Φ q(t) dt, (2)
0
where the variation vanishes at the endpoints and fulfills the nonholonomic constraints:

]
δq(t) = dLq(t) (e) δq(t), 0 = δq(0) = δq(T ), B q(t) · δq(t) = 0. (3)
The usual approach for Hamilton’s principle on Lie groups [3] together with the approach for the nonholo-
nomic Lagrange-d’Alembert principle [6] and the kinematic equation (1) yields the equations of motion
T T T
bT · M · v − DU(q) − DΦ(q) · λ − B(q) · µ,
M · v̇ = v (4a)
0 = Φ(q), (4b)
0 = B(q) · v, (4c)
where µ(t) ∈ Rm2 are additional Lagrange multipliers associated with the nonholonomic constraints.
2 The nonholonomic RATTLie time integration scheme
In order to construct the nonholonomic RATTLie time integration scheme, we use the approach of
variational integrators [7] and start with the variational principle (2). We replace q(t) by a discretized
version q d (t) defined by
t − t
k−1
q d (t) = qk−1 ◦ exp v^
k−1/2 , t ∈ [tk−1 , tk ),
h
where tk = t0 + kh are equidistant grid points with step size h = T /N . The vk−1/2 are defined in
such a way that q d is continuous. The function q d is differentiable for t ∈ (tk−1 , tk ) with constant
velocity vector vk−1/2 . Furthermore, we replace λ(t) by a discretized version λd (t) that we allow to be
discontinuous at tk and define λ± k = limτ →tk ±0 λ (τ ). By splitting the integral into several integrals over
d
(tk−1 , tk ), evaluating the integral over the kinetic energy and approximating the remaining integrals with
PN −
the trapezoidal rule, we arrive at the discrete variational principle 0 = δ k=1 Ld (qk−1 , qk , λ+ k , λk ) with
variations that fulfill a discretized version of (3) and with the discrete augmented Lagrangian
− h T h h h T h T
Ld (qk−1 , qk , λ+
k , λk ) = vk−1/2 ·M·vk−1/2 − U(qk−1 )− U(qk )− Φ(qn−1 ) ·λ+
n−1 − Φ(qn ) ·λ−
n.
2 2 2 2 2
208
The discrete variational principle leads to a form of discrete Euler-Lagrange equations:

− −
hµTk · B(qk ) = D2 Ld (qk−1 , qk , λ+ d +
k , λk ) + D1 L (qk , qk+1 , λk+1 , λk+1 ),
− −
0 = d3 Ld (qk−1 , qk , λ+
k , λk ), 0 = d4 Ld (qk−1 , qk , λ+
k , λk ).
The term hµTk · B(qk ) results from the fact that we only consider variations that satisfy the nonholonomic
constraint. We will apply a discrete Legendre transformation, where we write the impulses as M · vk ,
where vk is the velocity vector. The equations 0 = d3 Ld and 0 = d4 Ld are both equal to the constraint
equations 0 = Φ, so we replace one of them by the hidden constraints DΦ(q) · v = 0, which we obtain by
differentiating the constraint equation (4b). Plugging everything together, we arrive at the nonholonomic
RATTLie integration scheme:
qk = qk−1 ◦ exp(hv^
k−1/2 ),
h h T h T
M · vk−1 = T−T (−hvk−1/2 ) · M · vk−1/2 − fk−1 + DΦ(qk−1 ) · λ+
k−1 + B(qk−1 ) · µk−1 ,
2 2 2
0 = Φ(qk ),
h h T h T T
M · vk = T−T (hvk−1/2 ) · M · vk−1/2 + fk − DΦ(qk ) · λ−
k − B (qk ) · µk ,
2 2 2
0 = DΦ(qk ) · vk , 0 = B(qk ) · vk .
T
Here, fk = DU(qk ) + F (tk , qk ) is the sum of external, conservative and nonconservative forces. The
−
Lagrange multipliers λk = (λ+ k + λk )/2 and µk become unstable for fine step sizes h, so we have to
calculate corrected Lagrange multipliers by solving
 T 
 T T   corr  ck · M · vk + fk
v
M DΦ(qk ) B(qk ) v̇k  DΦ(q 
 k−1 ) − DΦ(qk+1 )
DΦ(qk ) 0 0  ·  λ corr 
k = · vk 
.
corr  2h 
B(qk ) 0 0 µ k B(qk−1 ) − B(qk+1 )
· vk
2h
3 The rolling disk
We consider an infinitely thin disk of radius 1 which rolls without slip on the plane P1,2 = {x ∈
R3 : [0, 0, 1] · x = 0}. We chose S3 × R3 3 (p, x) as configuration space, where x is the position of
the disk’s center point and p is the disk’s orientation as a unit quaternion. We want to describe both
velocity U and angular velocity Ω with respect to the body-fixed frame. Therefore, the configuration
space gets the structure of the semidirect product Lie group (S3 n R3 , ◦), similar to SE(3). For the
orientation p = [1, 0, 0, 0]T , the disk is supposed to lie flat in the plane P1,2 . Now we can calculate the
lowest point of the disk yL (p) with orientation p = [p0 , p1 , p2 , p3 ]T ∈ S3 with respect to the body-fixed
frame:
T q
yL (p) = p0 p2 − p1 p3 , p2 p3 − p0 p1 , 0 / (p21 + p22 )(p20 + p23 ).
p
In order for the disk to roll on the plane P1,2 it has to hold x3 = −2 (p21 + p22 )(p20 + p23 ) where x =
[x1 , x2 , x3 ]T . Since the disk should roll without slip, the overall velocity of the contact point
has to
vanish. From this, it follows C(p, x) · [Ω T , U T ]T = 0 with C(p, x) = x3 skw yL (p) , x3 I3 . If we
reformulate this constraint with respect to the inertial frame by applying the unit quaternion p to each
column of C(p, x), we get
B(p, x) Ω
0= · ,
DΦ(p, x) U
where Φ(p, x) = x23 /2−2(p21 +p22 )(p20 +p23 ). We can see that one condition is integrable and defines a holo-
nomic constraint, while the other two are not integrable and define two nonholonomic constraints. Plug-
ging everything together, we obtain the disk’s equations of motion (4) with q = (p, x), v = [Ω T , U T ]T ,
M = blkdiag(J, mI3 ) and U(q) = −mγ T · x, where γ ∈ R3 is the gravitational acceleration and J is the
inertia tensor of the disk with respect to the body-fixed frame. Due to the choice of the configuration
space and its Lie group structure, the inertia tensor J and the mass matrix M are constant diagonal
matrices.
209
1
Orientation
0.5 µ
101
λ
0
E
10−1 v
max. abs. error

−0.5
q
−1
0 0.5 1 1.5 2 10−3
6
10−5
Position / m
4
2 10−7
0 10−9
0 0.5 1 1.5 2 10−6 10−5 10−4 10−3
Time t / s Step size h / s
Figure 1. Orientation and position of the disk over Figure 2. Maxima of absolute errors in configuration,
time for h ≈ 10−6 s velocity, Lagrange multipliers and mechani-
cal energy E over step size h
4 Numerical tests
In order to test the nonholonomic RATTLie integration scheme, we have simulated the rolling disk with
mass m = 15 kg, inertia tensor J = m diag(0.25 m2 , 0.25 m2 , 0.5 m2 ) and gravity γ = [0, 0, −9.81 m/s2 ]T
p p T
from t0 = 0 s to T = 2 s. We chose the consistent initial values p = 0, 0, 1/2, 1/2 and x =
[0 m, 0 m, 1 m]T . The initial angular velocity was chosen to be Ω = [0 s−1 , 15 s−1 , 5 s−1 ]T and the initial
velocity U was chosen to be consistent. We have shown the evolution of orientation and position of the
disk in Figure 1.
In Figure 2, we have plotted the maxima of the absolute errors in the configuration q, velocity v, Lagrange
multipliers λ and µ as well as in the mechanical energy E = T +U over the step size in double logarithmic
scale. We have calculated the errors with respect to a numeric solution obtained by applying the same
integration scheme with very fine step size h = 2−20 s ≈ 10−6 s. We can see that if we scale the step size h
by a factor c, we can expect that the errors will be scaled by a factor c2 . This suggests that the scheme is,
like the original RATTLE [1], an integration scheme of second order [5]. As we expect from a variational
integrator [7], the nonholonomic RATTLie scheme does not show a systematic drift in mechanical energy.
References
[1] H. C. Andersen. Rattle: A “velocity” version of the shake algorithm for molecular dynamics calcula-
tions. Journal of Computational Physics 52.1 (1983), pp. 24–34.
[2] M. Arnold, A. Cardona, and O. Brüls. A Lie algebra approach to Lie group time integration of
constrained systems. In: Structure-Preserving Integrators in Nonlinear Structural Dynamics and
Flexible Multibody Dynamics. Ed. by P. Betsch. Springer, 2016, pp. 91–158.
[3] O. Brüls, M. Arnold, and A. Cardona. Two Lie group formulations for dynamic multibody systems
with large rotations. In: Proceedings of IDETC/MSNDC 2011, ASME 2011 International Design
Engineering Technical Conferences. Washington, USA, 2011.
[4] S.J. Ferraro, D. Iglesias-Ponte, and D.M. de Diego. Numerical and geometric aspects of the nonholo-
nomic SHAKE and RATTLE methods. Discrete Contin. Dyn. Syst (2009), pp. 220–229.
[5] E. Hairer and G. Wanner. Solving Ordinary Differential Equations II. Stiff and Differential-Algebraic
Problems. 2nd ed. Springer-Verlag Berlin Heidelberg, 2002.
[6] W.S. Koon and J.E. Marsden. The Hamiltonian and Lagrangian approaches to the dynamics of
nonholonomic systems. Reports on Mathematical Physics 40.1 (1997), pp. 21–62.
[7] J.E. Marsden and M. West. Discrete mechanics and variational integrators. Acta Numerica 10 (2001),
pp. 357–514.
210

A Brief Survey on Non-standard Constraints: Simula-

tion and Optimal Control
1
Dominik Kern ∗
and Ulrich J. Römer2
1 TU Chemnitz, Germany
2 KarlsruheInstitute of Technology (KIT), Germany
Abstract. In terms of simulation and control holonomic constraints are well documented and thus termed
standard. As non-standard constraints, we understand non-holonomic and unilateral constraints. We limit
this survey to mechanical systems with a finite number of degrees of freedom. The long-term behavior of non-
holonomic integrators as compared to structure-preserving integrators for holonomically constrained systems
is briefly discussed. Some recent research regarding the treatment of unilaterally constrained systems by event-
driven or time-stepping schemes for time integration and in the context of optimal control problems is outlined.
1 Introduction
Many technical systems are modeled as multibody systems which are comprised of rigid or elastic bodies,
joints, and force elements or forcing. We restrict the discussion to discrete systems, meaning elastic bodies
are presumed to be discretized by shape functions or finite elements and the system is represented by a
finite number of degrees of freedom. Joints are the typical pictorialization of constraints; for example they
constrain a body to move only along or around a certain axis. Generally, constraints may either account
for kinematic restrictions of a body, such as with redundant coordinates, or for interactions between
bodies and the environment. Such interactions may be described by an equation, such as a pivot, or by
an inequation, such as a non-penetration condition.
Without any claim for completeness, we report about some active research related to simulation and
control. We adopt the common classification of constraints into: holonomic or non-holonomic (e. g.
massless rigid connectors or rolling without sliding), scleronomic or rheonomic (e. g. static or moving
environment), and unilateral or bilateral (e. g. limiters or bearings). Bilateral holonomic, particularly
holonomic-scleronomic, constraints are the simplest type of constraints and their treatment has been well
studied and is covered in many textbooks [12, 32]. Therefore, we refer to them as standard. There are
some ongoing discussions regarding the remaining constraint types, non-holonomic and unilateral, which
is why we refer to them as non-standard.
2 Non-holonomic constraints
We need to recall the definition of a holonomic constraint to give meaning to the term non-holonomic.
Having a multibody system with generalized coordinates q in mind, a holonomic-scleronomic constraint
is given by an algebraic equation Ψ(q) = 0 on position level, whereas a holonomic-rheonomic constraint
would explicitly depend on time t. For further details on non-holonomic systems we refer to the classical
textbook of Neimark & Fufaev [25] and more recently Rabier & Rheinboldt and Bloch [5, 26]. A popular
formulation for the numerical solution of holonomically constrained systems was proposed by Gear, Gupta
and Leimkuhler [11]
∗ Corresponding author: dominik.kern@mb.tu-chemnitz.de ORCID iD: 0000-0002-1958-2982
211
q̇ = v − GT(q)η , (1a)
T
M(q)v̇ = f (q, v) − G (q)λ , (1b)
Ψ(q) = 0 , (1c)
G(q)v = 0 , (1d)
with mass matrix M(q), generalized forces f , Lagrange multipliers λ and η, and G = ∂Ψ/∂q. It
enforces the constraints on both, position and velocity level, and is also referred to as stabilized index-2
formulation. A combination of differential and algebraic equations like in the example given by eq. (1) is
called differential algebraic equation (DAE) or descriptor system.
Non-holonomic constraints are usually understood as equations between differentials of dq that cannot be
integrated to an equation on position level. A general criteria for non-integrability gives the Frobenius-
Theorem [14]. In mechanical systems, where most constraints appear as linear forms A(q)q̇ = 0, there
is a simple necessary condition. For a non-holonomic constraint the derivative of the corresponding row
in A must be a non-symmetric matrix.
From the point of view of numerics, the inclusion of non-holonomic constraints into eq. (1) is straight-
forward: they are added to the holonomic constraints on velocity level. This formulation works well
for moderate time spans. The robustness and the excellent long-time behavior of structure-preserving
integrators for unconstrained and holonomically constrained systems [13] suggest taking advantage of the
geometrical properties of non-holonomic systems. However, non-holonomic systems do not, in general,
preserve a symplectic structure. There is the open question whether energy conservation is the only
geometric property of non-holonomic systems. Consequently, an energy-consistent scheme has already
been proposed [4] to include both, holonomic and non-holonomic constraints. Recent research indicates
that the good long-term behavior of current non-holonomic integrators is due to the special properties of
popular benchmark systems and does not hold in general [22].
Lie-group methods [8, 24] and their successful application for the description of large rotations are not
considered further, since there is no hope to identify a Lie-group for any possible constraint.
The consideration of non-holonomic constraints in optimal control problems is also more challenging than
for holonomic ones. We restrict the discussion to feed-forward control, meaning we calculate a nominal
trajectory that is optimal in the sense of a given cost functional. These are constrained optimizations,
because the equations of motion, or more generally the system dynamics, are imposed as constraints on
the optimal control problem. For a general discussion, we refer to the comprehensive textbook [12] and
take a glance at concepts motivated by structure preservation. Both, the discretization-based approach
Discrete Mechanics and Optimal Control of Constrained systems (DMOCC) [19] and the functional-
based approach [15], are designed for holonomic constraints. Nevertheless, both allow different time
discretizations than variational integrators. Consequently, it will be interesting to find out how an
energy-consistent time discretization like the one proposed by Betsch [4] performs in this context.
3 Unilateral constraints
While bilateral constraints describe relationships between some or all state variables via equations, unilat-
eral constraints introduce one-sided limits on the admissible domain which result in inequations φ(q) ≥ 0
or inclusions. Consequently, solutions may be non-smooth or even discontinuous. In mechanics, typical
origins of unilateral constraints are collisions and friction. Several approaches exist to describe such sys-
tems, we refer to [6, 17] for further details. Numerical methods for approximate solutions have to deal
with the (piecewise) non-smooth dynamics and follow one of two basic philosophies: event-driven or time-
stepping, cf. figure 1. Event-driven schemes rely on algorithms for ODEs or DAEs to solve the smooth
Figure 1. Numerical approximation of time behavior of state qj : left smooth solution (variable time step); middle
non-smooth solution with event-driven approach (variable time step); right non-smooth solution with
time-stepping scheme (fixed time step).
212
parts of the solution trajectory [3], discrete events where inequalities become active are determined itera-
tively [29], a discrete state update is evaluated and the smooth solver is restarted. Time-stepping schemes
approximate the solution trajectory at fixed time steps where constraints are satisfied exactly (subject
only to numerical precision), but might be (slightly) violated in between [1, 6]. Event-driven schemes are
standard in most commercial software tools, since high-order methods can be used for the smooth parts
of the solution and good performance is achieved if there is a moderate number of events. They perform
badly if there is a high number of events or even accumulation points (Zeno behavior), since this leads to
very small time steps and the algorithm gets stuck. Furthermore, existence and uniqueness of a solution
of the ODE or DAE after an event is not guaranteed (cf. the Painlevé paradox) [31]. Time-stepping
schemes overcome these problems by “averaging” over the influence of all events during fixed time steps
[23]. Velocity jumps due to collisions are treated by a coefficient of restitution which acts as though all
collisions within a time step occur simultaneously [30]. They can be formulated as implicit time integra-
tion methods which require the solution of a linear or nonlinear complementary problem at every time
step [2].
In terms of structure-preservation, event-driven collision integrators follow naturally from Hamilton’s
principle, when the collision time is an additional variable [9]. Note, that this use of Hamilton’s principle
in equality form [9] does not allow trajectories to approach or to exit motions along the admissible
set boundary, i.e. contacts being opened or closed. In addition, the adjusted time steps may slightly
violate the conservation properties of variational integrators and thus accumulate errors. The extension of
classical analytical mechanics to perfect unilateral constraints, with possible persistent contact, requires a
reformulation of the variational principles of mechanics in terms of variational inequalities [18]. It is thus
natural to consider a nonsmooth, discrete Variational Principle for inequality constrained systems [16].
Since variational integrators are formulated in terms of momenta and avoid direct computation of forces
they are a convenient choice for time-stepping schemes. There is ongoing research in the development
of higher order methods to increase time step size and performance [28]. In the context of transient
simulations, the competition between event-driven and time-stepping schemes is not decided yet and
maybe never will.
For the optimal control problem (feed-forward), we distinguish two cases: firstly one only requires com-
pliance with the unilateral constraints, secondly there are collisions with the admissible boundary set.
The first case, pure compliance with unilateral constraints such as actuator saturation, is comparatively
simple and commonly accounted for by the Karush-Kuhn-Tucker conditions [7]. The second case, when
the events of interaction with the boundary of the admissible domain become subject of the optimization,
requires more sophistication since then the related dynamics are non-smooth and the events, their number
and their states, are additional unknown variables. As an example, consider a pendulum-on-cart-system
where the cart moves between two limiters and the collisions with these boundaries are to be utilized
for a swing-up of the pendulum. Since classical control theory presumes smooth dynamics, the direct
approach, not to say brute-force, splits this problem into two layers. The inner layer finds the optimal
solution between two collisions. The earlier collision is given by a fixed post-collision state and the later
one by a fixed pre-collision state, defining already the subsequent post-collision state. The outer layer is
used to find the optimal collision times and states [10]. A further outer layer may optimize the contact
mode sequence, which is an integer programming problem growing exponentially with the number of
contact constraints. Alternatively, time transformations [27] or contact-implicit trajectory optimization
[21] can be used to avoid this mixed integer optimal control problem. We further mention that there is
a close relation to the theory of switched systems [20].
References
[1] V. Acary and B. Brogliato. Numerical methods for nonsmooth dynamical systems: applications in
mechanics and electronics. Berlin: Springer, 2008.
[2] M. Anitescu. Optimization-based simulation of nonsmooth rigid multibody dynamics. Math Pro-
gram 105.1 (2006), pp. 113–143.
[3] M. Arnold and K. Frischmuth. Solving problems with unilateral constraints by DAE methods. Math
Comput Simulat 47.1 (1998), pp. 47–67.
[4] P. Betsch. A Unified Approach to the Energy-Consistent Numerical Integration of Nonholonomic
Mechanical Systems and Flexible Multibody Dynamics. GAMM-Mitteilungen 27.1 (2004), pp. 66–
87.
213
[5] A. M. Bloch. Nonholonomic mechanics. In: Nonholonomic mechanics and control. Vol. 24. New
York: Springer, 2003, pp. 207–276.
[6] B. Brogliato. Nonsmooth Mechanics. 2nd ed. London: Springer, 1999.
[7] R. Cottle and C. E. Lemke. Nonlinear programming. Vol. 9. SIAM-AMS Proceedings. American
Mathematical Soc., 1976.
[8] P. E. Crouch and R. Grossman. Numerical integration of ordinary differential equations on mani-
folds. J Nonlinear Sci 3.1 (1993), pp. 1–33.
[9] R. C. Fetecau, J. E. Marsden, M. Ortiz, and M. West. Nonsmooth Lagrangian mechanics and
variational collision integrators. SIAM J Appl Dyn Syst 2.3 (2003), pp. 381–416.
[10] K. Flaßkamp and S. Ober-Blöbaum. Variational formulation and optimal control of hybrid La-
grangian systems. In: P Int Conf HSCC. ACM. 2011, pp. 241–250.
[11] C. W. Gear, B. Leimkuhler, and G. K. Gupta. Automatic integration of Euler-Lagrange equations
with constraints. J Comput Appl Math 12 (1985), pp. 77–90.
[12] M. Gerdts. Optimal control of ODEs and DAEs. Berlin: De Gruyter, 2011.
[13] E. Hairer, C. Lubich, and G. Wanner. Geometric numerical integration: structure-preserving algo-
rithms for ordinary differential equations. Berlin: Springer, 2006.
[14] T. Hawkins. Frobenius, Cartan, and the problem of Pfaff. Arch Hist Exact Sci 59.4 (2005), pp. 381–
436.
[15] E. R. Johnson and T. D. Murphey. Scalable variational integrators for constrained mechanical
systems in generalized coordinates. IEEE T Robot 25.6 (2009), pp. 1249–1261.
[16] D. M. Kaufman and D. K. Pai. Geometric numerical integration of inequality constrained, nons-
mooth Hamiltonian systems. SIAM J Sci Comput 34.5 (2012), A2670–A2703.
[17] J. Kleinert and B. Simeon. Differential-algebraic equations and beyond: from smooth to nonsmooth
constrained dynamical systems. arXiv:1811.07658 (preprint) (2018).
[18] R. I. Leine, U. Aeberhard, and C. Glocker. Hamilton’s principle as variational inequality for me-
chanical systems with impact. J Nonlin Sci 19.6 (2009), p. 633.
[19] S. Leyendecker, S. Ober-Blöbaum, J. E. Marsden, and M. Ortiz. Discrete mechanics and optimal
control for constrained systems. Optim Contr Appl Met 31.6 (2010), pp. 505–528.
[20] D. Liberzon. Switching in systems and control. Boston: Birkhäuser, 2003.
[21] Z. Manchester, N. Doshi, R. J Wood, and S. Kuindersma. Contact-implicit trajectory optimization
using variational integrators. Int J Rob Res (2019).
[22] K. Modin and O. Verdier. What makes nonholonomic integrators work? arXiv:1707.04514 (preprint)
(2019).
[23] J. J. Moreau. Unilateral contact and dry friction in finite freedom dynamics. In: Nonsmooth Me-
chanics and Applications. Vienna: Springer, 1988, pp. 1–82.
[24] H. Munthe-Kaas. Runge-Kutta methods on Lie groups. BIT 38.1 (1998), pp. 92–111.
[25] J. I. Neimark and N. A. Fufaev. Dynamics of nonholonomic systems. American Mathematical
Society, 2004.
[26] P. J. Rabier and W. C. Rheinboldt. Nonholonomic Motion of Rigid Mechanical Systems from a
DAE Viewpoint. Philadelphia: SIAM, 2000.
[27] M. Ringkamp, S. Ober-Blöbaum, and S. Leyendecker. On the time transformation of mixed integer
optimal control problems using a consistent fixed integer control function. Math Program 161.1-2
(2017), pp. 551–581.
[28] T. Schindler and V. Acary. Timestepping schemes for nonsmooth dynamics based on discontinuous
Galerkin methods: definition and outlook. Math Comput Simulat 95 (2014), pp. 180–199.
[29] L. F. Shampine and S. Thompson. Event location for ordinary differential equations. Comput Math
Appl 39.5-6 (2000), pp. 43–54.
[30] B. Smith, D. M. Kaufman, E. Vouga, R. Tamstorf, and E. Grinspun. Reflections on simultaneous
impact. ACM T Graphic 31.4 (2012), pp. 1–12.
[31] D. E. Stewart. Rigid-body dynamics with friction and impact. SIAM Rev 42.1 (2000), pp. 3–39.
[32] G. Wanner and E. Hairer. Solving ordinary differential equations II. Berlin: Springer, 1996.
214

Comparison of timestepping schemes for non-smooth

mechanics using orbital convergence
1
Paschkowski, Manuela ∗
and Arnold, Martin1
1 Martin Luther University Halle-Wittenberg, Institute of Mathematics, 06099 Halle (Saale), Germany
Abstract. The dynamics of mechanical systems with non-smooth forces like friction, contacts and impacts
can be simulated by differential equations in combination with inequality constraints. However a con-
straint gets active, parts of the solutions characterising the velocity of the system deal with jumps, with
discontinuity points. Timestepping schemes are well-known possibilities to integrate such dynamical sys-
tems numerically. Their advantage is the avoided event detection such that a large number of disconti-
nuity points – especially accumulation points of velocity jumps – can be handled with higher computa-
tional efficiency. These schemes are always of integration order one with respect to discrete Lp -norms.
Hence, this classical tool can not be used to compare different algorithms. A more reasonable tool to
compare the approximation accuracy is the concept of orbital convergence. This new concept is stud-
ied for time-independent systems with invariant limit cycles using an experimental convergence analysis.
1 Non-Smooth Mechanics
In the following, the motion of a multibody system connected to joints or other mechanical elements like
springs or dampers is of interest. The generalised coordinates are represented by q : I ⊂ R → Rn and the
velocity by dq =: v : I → Rn . The degree of freedom of the motion can be reduced by involving interaction
forces like contacts or impacts effecting on the multibody system in form of additional inequalities
g(q(t)) ≥ 0, t ∈ I, g : Rn → Rm .
There are critical points t ∈ I where a positive constraint becomes zero. In these points summarised in
T ∗ := {t ∈ I : ∃ δ > 0, i = 1, . . . , m s.t. gi (q(τ )) > 0, ∀τ ∈ [t − δ, t), gi (q(t)) = 0}
v can be discontinuous and q non-differentiable. Combining classical variational principles with Lagrange
multiplier methods, the equations of motion of non-smooth mechanical problems have the form
M (q(t))dv(t) = f (q(t), v(t))dt + G(q)T (λ(t)dt + Λ(t)dη), (1)
0 ≤ λ(t) ⊥ g(q) ≥ 0, (2)
+ −
0 ≤ Λ(t) ⊥ G(q)(v + v ) ≥ 0, (3)
X
η= δs (4)
s∈T ∗ (q(t))
where M is the mass matrix, G = ∂g/∂q the constraint matrix, f the vector of the external forces, λ the
Lagrange multipliers representing the contact forces, Λ the Lagrange multipliers representing the impact
forces, ∈ [0, 1] the resolution number, dt the Lebesgue measure and δs the Dirac measure in time point
s. The formulation 0 ≤ a ⊥ b ≥ 0 is equivalent to a, b ≥ 0, aT b = 0 and is called complementarity problem.
A simple non-smooth mechanical system is the impact oscillator [2] in Figure 1 which motion is influenced
by a periodic force p and an obstacle in position σ. Of course, the oscillator can not penetrate the obstacle.
The inequality g(q) = σ −q ≥ 0 symbolises this condition. Everytime the oscillator and the obstacle get in
contact (g(q) = 0), an impact can be observed. This means, q gets non-differentiable and v discontinuous
(see Figure 2). The force λ induces that the inequality constraint is fulfilled for all following time points
and the impact force Λ represents the impact intensity.
∗ Corresponding author: manuela.paschkowski@mathematik.uni-halle.de
215
p
q
σ
Figure 1. Impact oscillator. Figure 2. Impact oscillator: Trajectories (σ = 0, p = sin(πt/8), = 0.6).
2 Non-Smooth Timestepping Schemes
There are two main groups of numerical methods to solve (1)-(4). The event-driven ones identify impact
points t ∈ T ∗ exact. Those methods are very accurate when dealing with few critical points, but get
inconsistent when the number increases or even an accumulation point of jumps can be observe like for
the impact oscillator. For this situation, the timestepping schemes have been developed. They define
a partition of the time interval t0 < t1 < . . . < tN = T with step sizes hi := ti − ti−1 which is never
changed during the whole integration process and therefore independent of the location of impacts. In
ti approximations qi ≈ q(ti ), vi ≈ v(ti ) are calculated in different ways. The first timestepping scheme
was developed in [3], uses the explicit Euler method between impacts and does not distinguish between
contact forces λ and impact forces Λ (see Algorithm 1).
Algorithm 1 Explicit Moreau-Jean timestepping scheme [3]

Require: time interval [t0 , T ], initial values q0 = q(0), v0 = v(0), i = 0, step size h = T /N
1: procedure MoreauJean(T, N, q0 , v0 , δ1 )
2: while ti+1 < T do
3: qi+1 = qi − hvi
4: M (qi )(vi+1 − vi ) = h(f (qi , vi ) + G(qi )T Λ) . Solve lines 6 and 7 simultaneously
5: 0 ≤ Λ ⊥ GJ 1 (qi+1 ) (qi )(vi+1 + vi ) ≥ 0
6: t = t + h, i = i + 1
7: end while
8: return q, v
9: end procedure
Algorithm 2 Forecasting trapezoidal rule [4]

Require: time interval [t0 , T ], initial values q0 = q(0), v0 = v(0), i = 0, step size h = T /N
1: procedure TrapezoidalRule(T, N, q0 , v0 , δ1 )
2: while ti+1 < T do
3: M (qi )(vi,1 − vi ) = h(f (qi , vi ) + GT (qi )λi )
4: 0 ≤ λi ⊥ g(qi ) ≥ 0
5: qi+1 = qi + h(vi + vi,1 )/2
6: M (qi+1 )(vi+1 − vi,1 ) = h (f (qi , vi ) + GT (qi )λi ) + (f (qi+1 , vi,1 ) + GT (qi+1 )λi1 ) /2 + GT (qi )Λi
7: 0 ≤ λi1 ⊥ g(qi+1 ) ≥ 0
8: 0 ≤ Λi ⊥ G(qi )(vi+1 + vi,1 ) ≥ 0
9: ti+1 = ti + h, i = i + 1
10: end while
11: return q, v
12: end procedure
Because the classical timestepping scheme in Algorithm 1 has only order of convergence one, in literature
a lot of numerical possibilities are studied to increase the accuracy or order. Different strategies are ex-
trapolation or step-size adaptation or retrospective bisections [1, 5]. They often deal with instabilities and
216
an inconsistent treatment of the non-smooth phases and are not satisfactory understood and analysed.
Based on this uncertain understanding of higher-order methods, in [4] a whole set of methods for the
consistent numerical treatment of (1)-(4) and their accuracy enhancement is developed based on discon-
tinuous Galerkin methods. These methods are special mortar methods which use discontinuous ansatz
and test functions in any situation not just if the situation demands it. One method of this algorithm
family is the forecasting trapezoidal rule in Algorithm 2. In smooth phases it uses the trapezoidal rule
of integration order two and distinguish between contact and impact forces.
3 Orbital Convergence and Test Implementations
Let (q, v) be the solution of (1)-(4) and (qih ), (vih ) ∈ Rn numerical approximations in t0 , t1 , t2 , . . . , tN ⊂ I
with constant step size h > 0. The global discretisation errors of a numerical scheme in tk are defined
as eqk := q(tk ) − qkh , evk := v(tk ) − vkh . Let eq = (eq0 , . . . , eqN )> , ev = (ev0 , . . . , evN )> and k.kN +1 be a
norm depending on the dimension N + 1 of e. A numerical scheme is called convergent of order p if
kekN +1 ∈ O(hp ). The approximation accuracy increases if p increases. Here, the L1 -norm is used
N
1 X
kekN +1,1 := kek k2 .
N +1
k=0
100 100
v - error in L1 -norm
q - error in L1 -norm
10-2 10-2
Moreau
Trapezoidal
10-4 10-4 O(h)
O(h2 )
10-6 10-6
10-3 10 -2
10 -1
10 0
10-3 10-2 10-1 100
step size h step size h
Figure 3. Classical convergence order for the impact oscillator.
In Figure 3, convergence order one for Algorithm 1 and 2 can be observed for the impact oscillator. It
makes no difference that Algorithm 2 uses a higher ordered method between impacts and a consistent
treatment of non-smooth forces. An increased computational effort for a better approximation of the
solution in the smooth phases is not reflected in the order of convergence at all. If one does not identify
the critical points in a certain accuracy like it would be typically for event-driven schemes, it may happen
that they are not noticed in the right time step. Figure 4 highlights visually the described phenomenon.
If there is a jump in the solution and it is noticed regardless of the step size first in the next time step,
the error remains almost constant. In such a point , the error is about as large as the jump. It may even
happen that a jump is registered at the correct interval with step size h and with h/2 not.
h = 0.06 h = 0.03 h = 0.015

0.6 0.6 0.6
error in v
error in v
error in v
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
0 5 10 0 5 10 0 5 10
t[s] t[s] t[s]
Figure 4. Absolute error in v: Impact oscillator.
In order to be able to compare different numerical methods, one needs another tool to judge their ap-
proximation quality. There are applications with a higher relevance of a good numerical approximation
217
of the whole solution set than of the approximation in certain time points. Noteworthy are invariant limit
cycles such as equilibrium points or periodic solutions. Therefore, the approximation quality of the whole
solution set, so-called orbits is now considered. It is important to mention, however, that the concept
is only plausible for applications with invariant limit cycles. Let x(t) = (q(t), v(t)) be the solution of
(1)-(4) with initial value
x(t0 ) = x0 . Then we call the set Γ(x0 , I) = {x(t), t ∈ I} orbit in x0 of I. The
discrete set Γh (I) = xhi = (qih , vih ), i = 0, 1, . . . , N is called numerical orbit of I. A numerical scheme
has orbital convergence order p if ρ γ h (I), γ(x0 , I) ∈ O(hp ) where
ρ(A, B) := sup inf ka − bk , A ⊂ Rm , B ∈ Rn
a∈A b∈B
is the one-sided Hausdorff-distance. In Figure 5, the orbit of the impact oscillator is plotted in black.
For both Algorithms 1 and 2, the numerical orbits for four different step sizes is added pointwise. They
approximate well the exact orbit; the ones belonging to Algorithm 2 better. Next, the numerical orbits
representing different step sizes are compared with the continuous one using the one-sided Hausdorff
distance. Figure 6 shows that we can observe orbital convergence order one for the Moreau-Jean scheme
and order two for the trapezoidal rule like we expect it.
(a) Moreau-Jean scheme (b) Trapezoidal rule

Figure 5. Impact oscillator orbits.
10-2
-4
10
ρ(Γh , Γ)
Moreau
10-6 Trapezoidal
O(h)
O(h2 )
10-8 -3 -2
10 10
step size h
Figure 6. Orbital convergence: Impact oscillator
References
[1] V. Acary. Higher order event capturing time-stepping schemes for nonsmooth multibody systems
with unilateral constraints and impacts. Applied Numerical Mathematics 62.10 (2012), pp. 1259–
1275.
[2] C. Budd, F. Dux, and A. Cliffe. The effect of frequency and clearance variations on single-degree-of-
freedom impact oscillators. Journal of sound and vibration 184(3) (1995), pp. 475–502.
[3] M. Jean and J.J. Moreau. Dynamics in the presence of unilateral contacts and dry friction: a nu-
merical approach. In: Unilateral Problems in Structural Analysis 2. Springer, 1987, pp. 151–196.
[4] T. Schindler and V. Acary. Timestepping schemes for nonsmooth dynamics based on discontinuous
Galerkin methods: definition and outlook. Mathematics and Computers in Simulation 95 (2014),
pp. 180–199.
[5] C.W. Studer. “Augmented time-stepping integration of non-smooth dynamical systems”. PhD thesis.
ETH Zurich, 2008.
218
MS 12: Computational Multiscale and Multiphysics Simulations of Microheterogeneous Materials

Investigation of the fractal properties of aerogels by

diusion-limited aggregation models
Rasul Abdusalamov
∗ 1, 1 2
Mikhail Itskov , Barbara Milow , Gudrun Reichenauer
3 and Ameya Rege
† 2
1 Department of Continuum Mechanics, RWTH Aachen University, Aachen, Germany

2 Institute of Materials Research, German Aerospace Center, Cologne, Germany
3 Division Energy Eciency, Bavarian Center for Applied Energy Research, Würzburg, Germany
Abstract. Aerogels are highly porous materials characterized by low thermal conductivities and sound ve-
locities. Most aerogels exhibit fractal morphology which strongly inuences their mechanical properties.
Diusion-limited aggregation (DLA) and diusion-limited cluster aggregation (DLCA) models [7] best de-
scribe the properties of silica aerogels [3, 6]. In our work, these models are implemented to characterize the
fractal features and mechanical properties of silica aerogels. To this end, the model results are rst validated
by comparison with the fractal exponents derived from experimental small angle X-ray scattering (SAXS)
data. Furthermore, the application of such aggregation models to describe the mechanical properties of these
aerogels is studied by importing the generated structure into a nite element analysis (FEA) tool.
Introduction
Silica aerogels belong to a class of highly open porous materials, those that are synthesized by the sol-gel
process. This process involves the assembly of smaller particles leading to the formation of clusters and
aggregates eventually forming a (percolating) backbone. Several studies in the literature have focused on
the thermal properties of silica aerogels given their super-insulating features. Due to their application
spectrum, there has also been a rise in investigations on their structural features and mechanical proper-
ties. Modeling and simulations have proven to be a very useful tool in understanding these properties in
many colloidal-like materials including silica aerogels. In this context, molecular dynamics simulations
[8] and coarse-grained modeling [2, 6] have been well exploited tools.
In this paper, an aggregation-based model approach is presented. Hasmy et al. [3, 4] investigated var-
ious aggregation-based algorithms and concluded that diusion-limited aggregation-based models best
describe silica aerogel microstructures. Since then, several studies have implemented such aggregation-
based models to describe the structural features of silica aerogels. In this work, highly porous structures
were formed, which were subsequently compared to experimental data from small angle X-ray scattering
(SAXS) measurements. As a result of the measurement, the mass fractal behavior is determined. Fur-
thermore, a representative volume element (RVE) was generated using the modeled structures to perform
a nite element (FE) simulation. Periodic boundary conditions (PBCs) were applied to the RVE. The
aim of the FE simulation was to investigate the bulk mechanical properties of the structures created by
means of the aggregation algorithm and compare them to experimental values of silica aerogels at given
porosity.
1 Generation of models using DLA and DLCA
DLA models were initially introduced in the 1980s by Witten and Sander [10]. In this case, the Brownian
motion of particles in the form of random walks is modeled on a regular cubic lattice. A total of NW
particles, often referred to as walkers, are initialized with an arbitrary position within this domain. These
particles then move in all directions according to the random walk process. In addition to the walkers,
∗ Corresponding author: abdusalamov@km.rwth-aachen.de ORCID iD: 0000-0003-4988-4794

† Corresponding author: ameya.rege@dlr.de ORCID iD: 0000-0001-9564-5482
219
(a) (b)
Figure 1. (a) Model silica aerogel aggregates generated by DLCA, (b) scanning electron microscopy image of
silica aerogel showing the aggregated morphology mimicked by DLCA (Reprinted by permission from
Springer Nature Customer Service Centre GmbH: [5], c Springer Science+Business Media, LLC 2008).
NS seeds are initialized. These are also located at random positions in the cube. If a walker is located
within a critical radius from the seed, it will be diused to it. Progressively, a tree of seeds similar
to a cluster-like structure emerges. The DLA model is extended if the seeds move in addition to the
walkers. The so extended model is known as diusion-limited cluster-cluster aggregation model (DLCA).
Furthermore, the dierent trees of the seeds can be linked to each other. Therefore, they also must be
within a critical radius. For the seeds as well as for the walkers, the width of the jumps can be specied
for the random walk in addition to this critical radius. In this work, the particles move on lattice points,
although this is also possible continuously. In addition, rotations of the clusters can also be implemented,
but are not considered in this work. According to Botet and Jullien [1], this does not aect the fractal
properties of the resulting clusters. In our case, a periodic structure is required, therefore the seeds as
well as the walkers are allowed to move beyond the cube borders and enter on the opposite side. The
code for the DLCA models was implemented in Matlab. Such a generated structure can be compared
with scanning electron microscopy (SEM) images showing real aerogel microstructures (see Figure 1).
2 Correlation of clusters to experimental SAXS measurements
In order to validate the correlation of the models to the experimental results, the mass fractal behavior
of the models must be determined. Accordingly, the structure generation is rst briey summarized. If a
cubic box with the side length L is considered and NW walkers are located in this region, the dimensionless
concentration or the volume fraction c can be determined. Assuming that the particles have a spherical
shape, c can be expressed by:
Vparticles 4π 3
c= with Vparticles = NW · R . (1)
L3 3
The edge length of the cube was L = 57 nm and in total 1137 seeds were created. For the spherical
particles, a diameter (2R = 3 nm) was chosen based on the experimentally evaluated silica aerogel particle
size. The evaluated c = 0.0868, where the porosity φ = (1 − c), which corresponds to 91.32%.
Figure 2a shows the experimental small angle scattering curve for a silica aerogel with a porosity of
about 92%. The above calculated φ is then in good correlation to the porosity of the silica aerogel as
determined from experiments. In the double-log plot of scattered intensity vs. scattering vector, the
curve is characterized by two regimes where the intensity is scaling with q : The q −df dependence at low
q-values indicates the scattering from clusters within the aerogel with a mass fractal dimension of df ,
which can directly be deduced from the slope in the double-log plot. At high q values, I scales with
q −4 indicating scattering from the particles forming the aggregates. The point where the two slopes are
intersecting can be converted in the characteristic particle size d [9].
220
To determine the mass fractal dimensions, the change of density as a function of distance r from a
reference point inside the aggregate was used. Accordingly, the mass m is expressed as
m ∝ rdf . (2)
where 0 < df < 3 is the fractal dimension of the aggregate.

The correlation of m versus r is presented in Figure 2b with df = 2.42 is obtained. This is in agreement
with the experimental SAXS data, where the fractal dimension of df = 2.32 ± 0.05 was obtained.
4
d~2/q
c0
10
d
I~q
f 103
3
10
I (a.u.)
2
10
log(M(r))
1
10
-4
I~q
0
102
10
DLCA model data
-1
d = 2.42
f
10
101
-1 0 1
10 10 10
-1
q (nm ) log(r)
Figure 2. (a) Log-log plot of scattered intensity versus scattering vector as measured by SAXS, (b) visualization
of the mass m(r) contained in concentric spheres with radius r plotted over r. The exponent of the
generated t is the fractal dimension df . In our case, df = 2.42.
3 Finite element simulation
For the FE simulations, the structures generated in Matlab were exported to LS-DYNA in the form
of nodes and elements. Here, the nodes represent the aerogel particle centers while the bonds between
the particles are modeled as beam elements with a circular cross-section. Beam elements were chosen
such that it accounts for the three modes of deformation that a bond undergoes: bending, torsion and
tension/compression. In addition, a frictional contact between the elements is dened. Since the structure
is completely periodic, the PBCs can be directly applied to the symmetry nodes. All nodes on the upper
and lower borders as well as the right and left borders are linked together. Hereby, the translational
and rotational degrees of freedom are coupled in the plane. With the help of dummy nodes, the desired
strains can be controlled. The resulting deformations for a 2-d structure can be seen in Figure 3, where
a compression of 10% is applied.
Figure 3. FE simulation of 10% compression applied to the generated 2-d aerogel structure (RVE).
221
4 Summary
In this paper, an aggregation-based model approach to characterize the fractal features and mechanical
properties of silica aerogels is presented. The mass fractal behavior is in agreement with typical values
derived from experimental small angle X-ray scattering data. For the study of their mechanical properties,
a nite element simulation of a 2-d aerogel structure is performed. As an outlook, 3-d silica aerogel
structures will be created and simulated to investigate their structural features and mechanical properties.
References
[1] R. Botet and R. Jullien. A theory of aggregating systems of particles : the clustering of clusters
process. Annales de Physique 13.3 (1988), pp. 153221.
[2] L. D Gelb. Simulating silica aerogels with a coarse-grained exible model and langevin dynamics.
The Journal of Physical Chemistry C 111.43 (2007), pp. 1579215802.
[3] A. Hasmy and E. Anglaret. Small-angle neutron-scattering investigation of long-range correlations
in silica aerogels: Simulations and experiments. Physical Review B 50.9 (1994).
[4] A. Hasmy, M. Foret, J. Pelous, and R. Jullien. Small-angle neutron-scattering investigation of short-
range correlations in fractal aerogels: Simulations and experiments. Physical Review B 48.13 (1993),
p. 9345.
[5] C. E. Kim, J. S. Yoon, and H. J. Hwang. Synthesis of nanoporous silica aerogel by ambient pressure
drying. Journal of Sol-Gel Science and Technology 49.1 (2008), p. 47.
[6] H-S. Ma, A. P. Roberts, J-H. Prévost, R. Jullien, and G. W. Scherer. Mechanical structureproperty
relationship of aerogels. Journal of Non-Crystalline Solids 277.2-3 (2000), pp. 127141.
[7] P. Meakin. Formation of fractal clusters and networks by irreversible diusion-limited aggregation.
Physical Review Letters51.13 (1983), pp. 11191122.
[8] S. P. Patil, A. Rege, Sagardas, M. Itskov, and B. Markert. Mechanics of Nanostructured Porous
Silica Aerogel Resulting from Molecular Dynamics Simulations. The Journal of Physical Chemistry
B 121.22 (2017), pp. 56605668.
[9] G. Reichenauer. Structural Characterization of Aerogels. In: Aerogels Handbook Advances in Sol-Gel
Derived Materials and Technologies.Ed. by M. A. Aegerter, N. Leventis, and M. M. Koebel. Springer
Science & Business Media, Mar. 2011, pp. 449498. isbn: 978-1-4419-7477-8. doi: 10.1007/978-
1-4419-7589-8_21.
[10] T. A. Witten and L. M. Sander. Diusion-limited aggregation. Chemistry Letters 27.9 (1983),
pp. 12111214.
222

Simulation of Dual-Phase Steel Microstructures Using

Optimized Arrangements of Finite Cells
∗
Yannick F. Fangye, Niklas Miska and Daniel Balzani
Ruhr-Universität Bochum, Lehrstuhl für Kontinuumsmechanik, Universitätstraße 150, 44801 Bochum, Germany
Abstract. In this contribution, the finite cell method is applied for the simulation of the microstructure of
dual-phase steel. As the geometry is typically given as voxel data, an alternative approach for the subcell
decomposition is presented, which exploits such data structures. The efficiency and accuracy of this new
approach is compared to the classical octree decompostion in two examples.
Introduction
The microstructure of dual-phase (DP) steel consists of a ferrite matrix with martensite inclusions.
For numerical simulations, this microstructure must be efficiently discretized. To this end, in [1, 2]
statistically similar RVEs (SSRVEs) were constructed, which have a lower complexity compared to the
real microstructure but still exhibit a similar mechanical behavior. The finite cell method (FCM) [3, 6]
facilitates the use of a regular finite element mesh and includes the microscopic material decomposition in
terms of subcells for integration purposes only. This approach is especially favorable for microstructures
given as voxel-based data, since the voxels can directly be considered to construct the integration mesh.
1 An optimal subcell decomposition for voxel-based geometry data
The FCM transfers the geometry representation from the finite element mesh to the integration level. It
is based on the use of a regular FE grid mesh with a decomposed integration, for which elements with
more than one material phases are subdivided into subcells. Each subcell consist then of a single phase.
Then, the integration of an element is given as
nc
X nc Z
X
e c
k = k = B T (ξξ (rr ))CcB (ξξ (rr )) det J e det J c dV , (1)
c=1 c=1 Bc
where k c denotes the stiffness matrix contribution of a single subcell and nc the number of subcells in the
element. The jacobian matrices J e and J c are respectively related to the mapping functions between the
local element (ξξ ) and the global coordinate system as well as the subcell (rr ) and the element coordinate
system.
In [3], a classic octree algorithm is used to construct the subcells, where the domain of interest is recur-
sively subdivided into eight equal regions. Thereby the phase fraction of the different materials varies
depending on the subdivision level. For voxel-based data, we propose to construct subcells that exactly
capture the material boundaries to overcome this issue (see also [5]). For this purpose a loop over all
spatial directions is carried out, wherein the voxels with the same material are successively merged in
order to build the largest possible subcells. To this end, all possible variations of orders of directions are
evaluated and the one with the least number of subcells is considered. In Fig. 1a an example of a simple
voxel-based geometry is shown, which is discretized using four finite elements. The resulting subcells for
both decomposition approaches, the octree split and the proposed algorithm, are depicted in Fig. 1b and
Fig. 1c respectively. For the octree decomposition, a relatively high subdivision level and therefore a high
∗ Corresponding author: daniel.balzani@rub.de
223
number of subcells is required to exactly capture the material boundaries. This may induce considerable
assembling effort. In contrast to that, the proposed approach constructs subcells that exactly coincide
with the borders of the voxels with considerably less subcells. However, it is possible that very narrow
subcells arise from this. Thus, the impact of these irregular subcells on the integration is investigated in
the following section. For this purpose, we introduce the subcell irregularity, defined as the ratio of the
longest to the smallest subcell edge, as characterizing measure.
a) b) c)
Figure 1. Illustration of a the considered voxel-based geometry, b resulting subcells from the octree method and c result-
ings subcells from the proposed optimal decomposition.
2 Numericals examples
In the first example the proposed subcell decomposition approach is compared with the octree method
in the context of the FCM. We consider a uniaxial tension test of a cube which is discretized with only
one 27-noded hexahedral finite element. The cube consists of 24 voxels in each spatial direction, in
which a martensite phase, described by an 11×11×11 voxel sized subcube, is positioned. The remaining
voxels belong to the ferrite phase. For the individual phases a finite J2 plasticity model (with isotropic
hardening) following [4] is used. The corresponding integration meshes are shown in Fig. 2a for the octree
method of level six with 5923 subcells and Fig. 2b for the proposed approach with only four subcells.
The results of the simulations for both approaches are depicted in Fig. 2c and Fig. 2d. It is observed
600 600
Optimal Decomposition
Octree
500
580
400
P max [MPa]
560
P x [MPa]
300 Optimal Decomposition

Octree-L1 540
200 Octree-L2
Octree-L3
Octree-L4 520
100
Octree-L5
Octree-L6
0 500
0 0.01 0.02 0.03 0.04 0.05 0 1 2 3 4 5 6 7 8 9
a) b) c) ∆lx /l0
d) nL
Figure 2. a Subcells for the octree of level six (5923), b subcells for the optimal decomposition approach, c homogenized
first Piola–Kirchhof stresses vs. relative elongation and d maximal macroscopic stress vs. subdivision level for
the octree method.
that the choice of the appropriate octree level is not straightforward although the results of the octree
split tend to the one of the optimal decomposition approach with increasing number of subcells.
For the next example, a more complicated microstructure is used, namely the SSVRE from [2]. Based on
a geometry representation with a resolution of 60×60×60 voxels, a mesh of 12×12×12 hexahedral finite
elements with cubic shape functions is considered. In addition to the previously introduced integration
methods, some variations of them are examined. In order to reduce the number of subcells arising from
the octree (T) decomposition within an element, the resulting subcells are, if possible, subsequently
merged (octree-merged (T-M) approach). As mentioned in Sec. 1, the optimal decomposition (OD)
may generate very narrow subcells. To avoid this, the octree-optimal decomposition (T-OD) method is
introduced, where the OD is performed within the finest octree level. The T-OD-M approach is similar to
the T-OD method and tries to merge all subcells arising from the T-OD split within an element. Fig. 3a
shows the subcells for the optimal decomposition approach. In Fig. 3b it is shown that the macroscopic
224
490 350
OD OD
T 300 T
485 T-M T-M
T-OD 250 T-OD
480 T-OD-M T-OD-M
Irregularity [-]
P max [MPa]
200
475 150
100
470
50
465
0
460 -50
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Level [-] Level [-]
b) c)
1e+07 40
OD OD
T 35 T
T-M T-M
Number of Subcells [-]

1e+06 T-OD 30 T-OD
T-OD-M T-OD-M
25
time [h]
100000 20
15
10000 10
5
1000 0
a) 0 1 2 3 4 5 6 0 1 2 3 4 5 6
Level [-] Level [-]
d) e)
Figure 3. a Subcells for the optimal decomposition approach, b maximal macroscopic stress, c irregularity, d number of
subcells and e computing time.
material responses of methods T and T-M are similar, since they have the same material boundaries and
only differ in the number of subcells and therefore consequently in the computing time as it can be seen
in Fig. 3d and Fig. 3e. Between T-OD and T-OD-M a similar behavior to T and T-M can be observed.
While the stress for T and T-M is changing depending on the subdivision level of the subcell and tend
with increasing subdivision level to the value of OD, the one of T-OD and T-OD-M remains constant
and has the same value as that for OD. The main advantage of the (-M) approaches is a reduction in the
number of subcells as compared to the preceding approaches. However, also an increase of the irregularity
of the subcells can be observed for the (-M) methods with increasing subdivision level, as depicted in Fig.
3c. In summary, the optimal decomposition method provides the best compromise in terms of accuracy,
computing time and irregularity.
Acknowledgments
The authors greatly appreciate financial funding by the German Science Foundation (Deutsche Forschungs-
gemeinschaft, DFG) as part of the priority program 1648, project BA2823/11-2.
References
[1] D. Balzani, L. Scheunemann, and D. Brands. Construction of two- and three-dimensional statistically
similar RVEs for coupled micro-macro simulations. Computational Mechanics 54 (2014), pp. 1269–
1284.
[2] D. Brands, D. Balzani, L. Scheunemann, J. Schröder, H. Richter, and D. Raabe. Computational
modeling of dual-phase steels based on representative three-dimensional microstructures obtained
from EBSD data. Archive of Applied Mechanics 86 (2016), pp. 575–598.
[3] A. Düster, J. Parvizian, Z. Yang, and E. Rank. The finite cell method for three-dimensional problems
of solid mechanics. Computer Methods in Applied Mechanics and Engineering 197 (2008), pp. 3768–
3782.
[4] S Klinkel. “Theorie und Numerik eines Volumen-Schalen-Elementes bei finiten elastischen und plas-
tischen Verzerrungen”. PhD thesis. Universitat Fridericiana zu Karlsruhe, 2000.
[5] N. Miska and D. Balzani. Quantification of uncertain macroscopic material properties resulting from
variations of microstructure morphology based on statistically similar volume elements: application
to dual-phase steel microstructures. Computational Mechanics (2019).
[6] D. Schillinger and M. Ruess. The Finite Cell Method: A Review in the Context of Higher-Order
Structural Analysis of CAD and Image-Based Geometric Models. Archive of Computational Methods
in Engineering 22 (2015), pp. 391–455.
225
226

Modeling interpenetrating metal matrix composites

using a fast Fourier transform formulation
1
Dominik Horny ∗
and Katrin Schulz1
1 Institute
for Applied Materials - Computational Materials Science (IAM-CMS), Karlsruhe Institute of Technology, Kaiserstr.12,
76131 Karlsruhe, Germany
Abstract. Just recently, a novel interpenetrating composite has been experimentally generated based on highly
porous ceramic foam preform. The heterogeneous microstructure of this interpenetrating Metal matrix com-
posite (MMC) is modeled in order to investigate the deformation behavior and basic failure mechanisms. We
use a fast Fourier transform (FFT) formulation to simulate high resolution grids of the delicate microstructre
with reasonable computational effort. Subjected to external loading, we investigate the mechanical behavior of
the material and the interactions on the microscale. Homogenized properties are derived from the simulations
and compared to experimental investigations of the material.
Introduction
Metal matrix composites with an interpenetrating microstructure made from Al2 O3 ceramic foam infil-
trated with AlSi12 cast alloy have been investigated. This class of materials is very promising since it
perfectly combines both good mechanical and thermal properties and has a notable lightweight potential
as well. Therefore, it can be used in the mobile sector e.g. to increase the efficiency of combustion engines,
thus reduce CO2 -emissions [4]. The Simulation with a high resolution grid and a reasonable computing
time requires more efficient methods than the commonly used finite Element method (FEM) due to the
complex and highly heterogeneous microstructure [8]. Here, the fast Fourier transform (FFT) method
provides an efficient tool to analyze the complex materials interaction [3]. We present a formulation
based on the spectral spectral solver framework of DAMASK (Düsseldorf Advanced Material Simulation
Kit) [5] to perform the FFT-based microstructure simulations and discuss the results with respect to the
homogeneous representation by an representative volume element (RVE) approach.
1 Fast Fourier transform based modeling
In order to describe the change of material point positions x in the reference configuration to y in a de-
formed configuration due to external load, a deformation map χ(x) is introduced. This deformation map
can be expressed as the sum of a homogeneous deformation gradient F and a superimposed deformation
fluctuation field w̃ with χ(x) = Fx + w̃ where the periodicity condition is constrained by w̃− = w̃+
on corresponding surfaces. Since the total deformation gradient is given by F = Gradχ it can also be
decomposed into homogeneous F and fluctuating part F̃ reading F = F + F̃.
The material response to external load is given by an energy density functional W: P(x) = δW/δF(x).
The direct variational formulation of the static equilibrium is obtained by minimizing W over all defor-
mation fields and can be expressed in real and Fourier space as min W = Div P(x) = F −1 [P(k)ik] = 0,
χ
with the frequency vector in Fourier space k and the Fourier transform operator F. It is equivalent to
finding the root of the residual body force field R[χ(k)] = P(k)ik = 0 with the preconditioning operator
R. Since this is an ill-conditioned PDE, a reference problem with a homogeneous linear reference material
is introduced. The well-conditioned PDE then reads
Rdirect [F(k)] := Γ(k)P(k) = 0 (1)

∗ Corresponding author: dominik.horny@kit.edu
227
with the Gamma operator Γ. More detailed explanation can be found in [3]. Discretizing the direct
variational formulation of the static equilibrium with a collocation-based approach at the grid points
leads to the so called "basic"-scheme:
"( #
−1 Γ(k)P(k) if k 6= 0
Rbasic [F(x)] := F . (2)
∆FBC if k = 0
The equation system (2) can be solved iteratively by e.g. non-linear GMRES solver or other non-linear
numerical solvers provided by PETSc library. The convergence criterion for the solution is defined by the
root mean square (RMS) of the divergence of stress introducing the equilibrium tolerance εeq and the
unit of length m:
RMS(DivP(x))
εeq × m ≥ . (3)
kPk2
For a more detailed explanation see references [3, 5, 6].
2 Effective elastic properties
We investigate a binarized micro computer tomography (µ-CT) reconstruction of an Al2 O3 foam specimen
with an edge length of around 1.9 mm (see Fig.1). The determined microstructural characteristics of the
foam are given in Tab.1.
Table 1. Microstructural characteristics

voxel size = 2.663 µm3
porosity = 69.6 %
mean pore radius = 14.39 µm
median pore radius = 10.01 µm
Figure 1. µ-CT scan of the Al2 O3 foam.

white = ceramic, black = pore
We choose different sizes of representative volume elements (RVEs) with edge lengths from 10-100 voxel
at 15 randomly chosen point inside the specimen to investigate the mechanical properties of the foam
and the MMC. The raw materials properties as well as properties of the ceramic foam and the MMC
were determined experimentally and are listed in Tab.2.
3 0 .4
1 8 0
1 6 0
Table 2. Experimentally determined R V E - s iz e [v o x e l]
Y o u n g 's m o d u lu s [ G P a ]
1 4 0 1 0 ³
material properties 1 2 0 M M C 2 0 ³
3 0 ³
Material E [GPa] µ 1 0 0 c e r a m ic fr a c tio n
4 0 ³
5 0 ³
Al2 O3 350 0.23 8 0 o f to ta l s c a n
6 0 ³
AlSi12 77.8 0.32 6 0 7 0

8 0
³
³
foam 31 0.29 4 0 fo a m
9 0
1 0
³
0 ³
MMC 127 0.27 2 0
0
0 5 1 0 1 5 2 0 2 5 3 0 3 5 4 0 4 5 5 0
c e r a m ic fr a c tio n [% ]
Figure 2. Simulation Results of RVE study on Effective

Young’s modulus.
228
For each RVE, six independent load cases (pressure in x-, y- and z-direction as well as shear in xy-, xz-
and yz-direction) were applied to determine the stiffness tensor C of the foam/MMC. Then, the Effective
Properties of C have been derived as described in [7]. The results of a comparison of the effective Young’s
modulus dependent on the ceramic volume fraction measured for the RVEs, respectively, is shown in
Fig.2.
3 Damage behavior
Stress distributions resulting from the elastic simulations of the MMC exemplify that stress peaks occur
especially in very thin ceramic rods. It is reasonable to assume that damage will first occur at these
points of stress concentration. Therefore, we analyze the damage behavior for a simplified benchmark
structure of a single ceramic rod embedded into an aluminum matrix, see Fig3.
Figure 3. Phase field damage evolution of an Al2 O3 rod embedded in AlSi12 under tensile load.
The damage and its evolution is prescribed by a phase field damage model introducing the scalar non-
local phase field damage variable ϕ varying between ϕ = 1 (undamaged) and ϕ = 0 (fully damaged).
The damage evolution is formulated by the classical Ginzburg-Landau equation [1] as a relaxation of the
overall free energy:

G
ϕ̇ = −M 2ϕSE − − ∇Gl0 ∇ϕ (4)
l0
Here M is the damage mobility, S the 2nd Piola Kirchhoff stress, E the Green-Lagrange strain, G the
fracture surface energy and l0 the characteristic length. A brittle damage is used which is driven by
the release of stored elastic energy ψel = 21 ϕ2 S · E which has to overcome the surface tension resistance
ψsurf = 21 Gl0 |∇ϕ|2 + lG0 (1 − ϕ) as described in detail in [2] and [5].
4 Discussion
Comparing the effective elastic properties given in Fig.2 and Tab.2 shows very good agreement of the
simulation results and the experimental values. Evaluating the FFT modeling techniques for the highly
heterogeneous MMC microstructures obtained by µ-CT scans, we derive reasonable elastic material prop-
erties and a robust as well as efficient computational method. Furthermore, the dominating role of the
ceramic fraction on the effective properties and the minor influence of the RVE-size itself could be shown
during the analysis. In conclusion a small RVE with the same ceramic fraction as the entire material is
sufficient to extract effective elastic Properties of the MMC. However, this holds true just to a limited
extent for the ceramic foam structure.
Damage was modeled for a benchmark problem and the evolution from undamaged to fully damaged
material has been shown as a first evaluation of a phase field approach coupled with the FFT formula-
tion. In further steps, the knowledge gained from this test structure will be used to model damage in the
significantly more complex interpenetrating microstructures.
229
Acknowledgments
The financial support for this work in the context of the DFG research project SCHU 3074/1 as well as
the support by the European Social Fund and the State Baden-Württemberg is gratefully acknowledged.
References
[1] S. M. Allen and J. W. Cahn. A microscopic theory for antiphase boundary motion and its application
to antiphase domain coarsening. Acta Metallurgica 27 (1979), pp. 1085–1095.
[2] M. Diehl, M. Wicke, F. Roters, P. Shanthraj, A. Brueckner-Foit, and D. Raabe. Coupled Crystal
Plasticity–Phase Field Fracture Simulation Study on Damage Evolution Around a Void: Pore Shape
Versus Crystallographic Orientation. Jom 69 (2017), pp. 872–878.
[3] P. Eisenlohr, M. Diehl, R. A. Lebensohn, and F. Roters. A spectral method solution to crystal
elasto-viscoplasticity at finite strains. International Journal of Plasticity 46 (2013), pp. 37–53.
[4] Daniel B. Miracle. Metal matrix composites - From science to technological significance. Composites
Science and Technology 65 (2005), pp. 2526–2540.
[5] F. Roters, M. Diehl, P. Shanthraj, P. Eisenlohr, C. Reuber, S. L. Wong, T. Maiti, A. Ebrahimi, T.
Hochrainer, H. O. Fabritius, S. Nikolov, M. Friák, N. Fujita, N. Grilli, K. G. F. Janssens, N. Jia,
P. J. J. Kok, D. Ma, F. Meier, E. Werner, M. Stricker, D. Weygand, and D. Raabe. DAMASK – The
Düsseldorf Advanced Material Simulation Kit for modeling multi-physics crystal plasticity, thermal,
and damage phenomena from the single crystal up to the component scale. Computational Materials
Science 158 (2019), pp. 420–478.
[6] P. Shanthraj, M. Diehl, P. Eisenlohr, F. Roters, and D. Raabe. Spectral Solvers for Crystal Plasticity
and Multi-physics Simulations. In: Handbook of Mechanics of Materials. Ed. by C. H. Hsueh and
et.al. Singapore: Springer, 2019, pp. 1–25.
[7] C. Soyarslan, S. Bargmann, M. Pradas, and J. Weissmüller. 3D stochastic bicontinuous microstruc-
tures: Generation, topology and elasticity - Supplementary. Acta Materialia 149 (2018), pp. 326–
340.
[8] J. Zeman, T. W. J. de Geus, J. Vondřejc, R. H. J. Peerlings, and M. G. D. Geers. A finite ele-
ment perspective on nonlinear FFT-based micromechanical simulations. International Journal for
Numerical Methods in Engineering 111 (2017), pp. 903–926.
230

Can the constituent stress-strain laws of composite ma-

terials be determined uniquely from integral measure-
ments?
∗
Fabian Sewerin
Institute of Solid Mechanics, Technische Universität Braunschweig, Langer Kamp 8, 38106 Braunschweig
Abstract.
Naturally occurring composite materials can serve as a blueprint for engineering materials. However, the
transfer of design principles is often complicated by the absence of detailed information on the material
behaviour of the individual constituents. Conversely, progress in imaging and non-destructive measurement
techniques has rendered the intra-sample arrangement of constituents accessible. In this contribution, we
enquire whether it is possible to conclude on the stress-strain laws of constituent materials based on force-strain
measurements on a composite sample whose geometrical structure is known. Since a priori intuition about
the functional forms of these stress-strain laws can be difficult to acquire, a discretization in terms of global
shape functions is explored. Starting from a candidate set of standard material tests, we propose the adoption
of a criterion for local qualitative identifiability that is based on the sensitivity matrix of model predictions.
Apart from an initial estimate, the analysis for local identifiability does not involve any experimental data
and may hence be used in the design process for an experimental campaign. As an example, we analyze
combinations of uni- and equibiaxial tension tests on a hyperelastic composite sheet for identifiability. The
force-strain response of the sheet is almost completely controlled by the volume ratio of two constituents.
Introduction
Composite materials are made from several constituents that are arranged and interconnected in a partic-
ular way. In engineering applications, composites are commonly designed with the objective to unify the
material properties of disparate constituents and create a hybrid that is more amenable to the intended
deployment than the constituents alone. Conversely, natural composite materials such as skeletal mus-
cle tissue or plant parenchyma can serve as a blueprint for artificially engineered composites. However,
in many cases, our understanding of the material behaviour of the natural constituents is limited and
experimental efforts to elucidate the mechanisms by which natural composites absorb and redistribute
mechanical loads have been challenging. For example, the isolation of individual constituents without
impairment can be extremely difficult if the geometrical arrangement of the constituents is intricate or
covers several length scales. At the same time, Oskay and Fish [2] argue that the material behaviour of
pure constituents in bulk may differ from the material behaviour a constituent exerts as part of a com-
posite, for example, on account of interfacial effects [3]. On the other hand, the geometrical arrangement
of constituents within a composite material can often be reconstructed with the aid of non-destructive
measurement techniques. By consequence, we are led to ask whether the material behaviour of the con-
stituents inside a composite sample can be inferred from standard material tests given records of the
geometrical intra-sample arrangement. Assuming that the model description is exact and measurement
errors are absent, this question falls into the realm of qualitative identifiability [4].
In practice, constitutive laws such as stress-strain relations are frequently determined by estimating a
functional form and, subsequently, identifying the free parameters using a least-squares optimization
based on experimental measurements and corresponding model predictions. However, in the absence of
experimental measurements on the material response of isolated constituents, it may be difficult to gain
or substantiate an a priori intuition about the functional forms of the stress-strain relations. In order to
mitigate this lack of intuition, we explore the direct discretization of a stress-strain relation in terms of
∗ Corresponding author: f.sewerin@tu-braunschweig.de ORCID iD: 0000-0001-6588-7739
231
global shape functions here. Although this idea affords generality in the inference of unknown functions,
it implicitly involves an assumption on the class of materials that dictates the physical nature of the
unknown functions, for example, a strain energy density governing a hyperelastic constituent.
In view of practical deployment, we adopt a sensitivity-based criterion for local identifiability. It is argued
that this criterion may be checked iteratively in the course of a Gauss-Newton scheme for estimating the
coefficients of the representation in terms of global shape functions. The Gauss-Newton sequence requires
an initial estimate for the coefficients and we provide evidence indicating that local identifiability is
maintained throughout the scheme if it holds for the initial estimate. Based on a library of candidate
material tests, those combinations which render the material model locally identifiable are determined.
As an example, we consider a composite sheet made from a regular arrangement of two hyperelastic
constituents. The material behaviour of the sheet is mainly influenced by the volume ratio of both
constituents and, only to a negligible extent, by the particular intra-sample constituent distribution.
Considering uniaxial and equibiaxial tension tests as candidate material tests, the local identifiability
analysis indicates that a minimum experimental campaign may involve both material tests and two
samples with different constituent volume ratios.
Parameter identification and local identifiability
Since our focus lies on standard material tests, we restrict the consideration to tests i which yield integral
forces Fi = F̂i (ε) over a prescribed engineering strain range ε ∈ Ei . Concomitantly, suppose that a
physical model is available which returns a force-engineering strain relation Fi = Fi (ε; w), ε ∈ Ei , for
given material parameters w. These parameters are commonly determined as solution of the following
least squares problem
XZ 2
Fi (ε; w) − F̂i (ε) dε = Min! (1)
Ei w
i
In a Gauss-Newton scheme to solve the necessary conditions of optimality for Eq. (1), an estimate wn ,
n = 0, 1, . . ., is updated iteratively by an increment βn+1 ∆wn+1 with
G(wn )∆wn+1 = −r(wn ), (2)
where r(w) represents the gradient with respect to w of the objective function in Eq. (1) and the left
hand side G(w) is given by
XZ T
∂Fi (ε; w) ∂Fi (ε; w)
G(w) = dε. (3)
i Ei ∂w ∂w
If we conceive of the vector ∂Fi (ε; w)/∂w for given ε as an observation, then G(w) corresponds to
the autocorrelation matrix of observations over the range ε ∈ Ei . In fact, application of a quadrature
approximation to Eq. (3) with weights γj and quadrature points εj shows that G(w) can be related to
a snapshot matrix X(w) whose dominant modes are frequently analyzed with the aid of a singular value
decomposition (SVD).
G(w) = X(w)X(w)T (4)

√ ∂F1 (ε1 ; w) √ ∂F1 (ε2 ; w)
X(w) = γ1 , γ2 ,... (5)
∂w ∂w
The transpose of X(w) coincides with the sensitivity matrix of the predicted force-engineering strain
√ √
curves ( γ1 F1 (ε1 ; w), γ2 F1 (ε2 ; w), . . .) sampled at the quadrature points and scaled by the square
roots of the corresponding weights.
A necessary condition for the Gauss-Newton scheme to converge is that the left hand side matrix G(w)
is regular at all points wn visited along the iteration sequence. By Eq. (4), this implies that the rank
of X(wn ) coincides with the number of material parameters in w. For practical purposes, we adopt the
criterion that the normalized smallest eigenvalue of G(w) exceeds a small tolerance ,
min{λ1 , λ2 , . . .}
λ̄min = P > , (6)
i λi
232
where the eigenvalues λi = σi2 are computed as the squares of the singular values of X(w). Since the
sensitivity matrix X(w)T does not involve any experimental measurements (Eq. (5)), the regularity of
G(w) at the initial estimate w0 can be assessed before material tests are undertaken. This attains even
more significance as we observed, for all examples considered and synthetic measurements, that regularity
at subsequent iterates wn , n = 1, 2, . . ., never ceased as long as G(w) was regular at w0 . Commensurately,
Vajda et al. [4] pointed out that if the rank of X(w0 ) coincides with the number of unknown parameters
in w0 , then the model for Fi (εj ; w) is locally identifiable at w0 .
This suggests that a subset of material tests for which Fi (εj ; w) is locally identifiable can be determined
out of a library of candidate tests by selecting an initial estimate w0 and assessing the information content
of X(w0 ) with the aid of the SVD-based criterion in Eq. (6).
A hyperelastic composite sheet
In order to corroborate the viability of the rationale developed above, we exemplarily consider a thin
composite sheet made from two hyperelastic constituents that are arranged in a grid-like pattern (Fig-
ure 1(a)). The composite sheet can be subjected to simple and equibiaxial tension tests [1] that yield
records of the horizontal and/or vertical forces over the applied engineering strain. On an integral basis,
the material is controlled by the volume ratio α ∈ [0, 1] of both constituents. By varying the spatial
distribution of the constituent patches while keeping α constant, we observed that the force-engineering
strain curves in either material test are almost exclusively controlled by the volume ratio α and to a
negligible extent influenced by the particular spatial arrangement. In this analysis, the first constituent
was taken as Neo-Hookean, while the second one was endowed with an exponential strain energy density.
In the Introduction, we argued that estimates of the functional forms of the strain energy densities
W1 (I1 , I3 ) and W2 (I1 , I3 ) characterizing the constituents may not always be available. In order to analyze
local identifiability despite this lack of information, W1 (I1 , I3 ) and W2 (I1 , I3 ) are approximated in terms
of global shape functions Hi (I1 , I3 ), i = 1, . . . , nH ,
nH
X (j)
Wjh (I1 , I3 ) = wi Hi (I1 , I3 ), j = 1, 2. (7)
i=1
Here, I1 and I3 denote the first and third invariants, respectively, of the right Cauchy-Green deformation
tensor and, for simplicity, the strain energy densities are assumed to be independent of the second
invariant I2 . The shape functions can be chosen almost arbitrarily, albeit subject to a linear independence
condition on the subset of I1 -I3 -space which is expected to be accessed in the course of the material tests.
Furthermore, we choose the shape functions in such a way that the strain energy densities Wjh as well
as the second Piola-Kirchhoff stresses Shj vanish in an unstretched reference configuration. As starting
(j)
point for a local identifiability analysis, the parameters wi are set to
(j) (j)
w1 = 200 N/m2 and wi = 0 for i = 2, . . . , nH = 5, j = 1, 2 (8)
such that, for the present choice of shape functions (not listed), both W1h (I1 , I3 ) and W2h (I1 , I3 ) coincide
and are given by a Neo-Hookean material with moduli µ = 200 N/m2 and λ = 0 N/m2 .
Figure 1(b) depicts a heatmap of the smallest normalized eigenvalue λ̄min of G(w) for combinations of
uniaxial and equibiaxial tension tests on two composite samples with volume ratios α1 and α2 . If the
threshold in Eq. (6) is chosen slightly larger than the computing precision, for example, = 10−14 ,
then it appears that the material law in Eq. (7) is locally identifiable for any (α1 , α2 ) pair. Beyond
this positive conclusion, Figure 1(b) also provides some insight into the relative information content of
different combinations of material tests. For example, λ̄min increases with the difference between α1
and α2 , indicating that the conditioning of G(w) improves as the samples are more dissimilar. Further
analyses on our part indicate that, in the present example, at least four material tests which differ
either in terms of the sample constitution α or the type of test (uniaxial or equibiaxial) are required for
local identifiability and that the addition of a fifth test does not improve upon the best conditioning in
Figure 1(b) for α1 = 1 and α2 = 0. At the same time, not all combinations of four, five or six material
tests result in local identifiability.
For those combinations of material tests for which local identifiability was ascertained in Figure 1(b), we
further quantified the reidentification accuracy assuming that the true constituents possess a Neo-Hookean
and an exponential strain energy density, respectively (not shown).
233
Smallest normalized
(a) (b) eigenvalue
1 1e−09
3e−10
0.9 6e−11
2e−11
0.8 4e−12
1e−12
0.7
Simple/Equibiaxial α2
û1
0.6
0.5
0.4
û2
0.3
0.2
2
0.1
5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Simple/Equibiaxial α1
1
Figure 1. (a) Schematic illustration of a composite sheet made from patches of two different hyperelastic con-
stituents in equibiaxial tension. The dotted lines indicate the unstretched reference configuration. (b)
Heatmap of the smallest normalized eigenvalue λ̄min for an experimental campaign involving one uniaxial
and one equibiaxial tension test on composite samples with two different volume ratios α1 6= α2 .
Conclusions
In order to support elucidating the material behaviour of constituents in composite materials, we ad-
dressed the question as to whether the constituent strain energy densities of hyperelastic composites
can be inferred from integral force-engineering strain measurements if information on the intra-sample
geometry is available. In the absence of a priori intuition about the functional form of these strain
energy densities, a direct discretization in terms of global shape functions was explored. Slightly reinter-
preting the governing question, we proposed to consider a library of candidate material tests and select
those combinations for which the sensitivity matrix of model predictions with respect to the material
parameters obeys a rank criterion. Apart from an initial estimate of the parameters, this criterion does
not yet involve any experimental data and may be employed to rationalize an experimental campaign
encompassing a minimum of material tests. The sensitivity-based criterion for local identifiability can
also be deployed iteratively in the course of a Gauss-Newton scheme for solving the parameter estima-
tion problem, but we observed that local identifiability at the initial estimate was sufficient to maintain
identifiability along the Gauss-Newton sequence.
Acknowledgments
I gratefully acknowledge Abimathi Siva Subramanian who assisted me in setting up the test case reported
here during a DAAD WISE internship in summer 2018.
References
[1] R. W. Ogden, G. Saccomandi, and I. Sgura. Fitting hyperelastic models to experimental data.
Computational Mechanics 34.6 (2004), pp. 484–502.
[2] C. Oskay and J. Fish. On calibration and validation of eigendeformation-based multiscale models for
failure analysis of heterogeneous systems. Computational Mechanics 42.2 (2008), pp. 181–195.
[3] U. Schmidt, J. Mergheim, and P. Steinmann. Multiscale parameter identification. International Jour-
nal for Multiscale Computational Engineering 10.4 (2012), pp. 327–342.
[4] S. Vajda, H. Rabitz, E. Walter, and Y. Lecourtier. Qualitative and quantitative identifiability analysis
of nonlinear chemical kinetic models. Chemical Engineering Communications 83.1 (1989), pp. 191–
219.
234

Numerical thermo-elasto-plastic analysis of residual stresses

on different scales during cooling of hot formed parts
1
S. Uebing ∗
, L. Scheunemann1 , D. Brands1 and J. Schröder1
1 Insitute of Mechanics, Civil Engineering, University Duisburg-Essen, Universitätsstr. 15, 45141 Essen, Germany
Abstract. The targeted application of residual stress states is an important tool in order to improve component
properties such as strength or durability. Therefore, in current research more and more attention is paid to the
numerical analysis on different scales, i.e. macroscale, mesoscale and microscale. Initiated by the hot forming
process of a cylindrical specimen, a microscopic phase transformation occurs which is directly linked to the
evolution of mesoscopic strains and related residual stresses. Hence, in this contribution two models are built,
one describing macroscopic stresses due to cooling and one describing the individual steps from transformation
to residual stress state on the lower scales.
Introduction
The experimental and numerical investigation of hot bulk forming processes is a useful tool to improve
the component properties, e.g. extension of lifetime. Hot forming processes offer the opportunity to
adjust some process parameters, i.e. deformation state, temperature profile or cooling media, in order to
influence the distribution of residual stresses. Therefore, underlying experiments and related macroscopic
numerical simulations of a hot formed cylindrical specimen made of Cr-alloyed steel 1.3505 (DIN 100Cr6)
are done at the Institute of Forming Technology and Machines (IFUM) at Leibniz University (LU)
Hannover. Detailed information regarding the procedure as well as resulting macroscopic residual stress
states can be found in [1]. Following [3], those stresses can be characterized as residual stresses of first
type. Residual stresses of second and third type refer to the lower scales, i.e. polycrystal and grain level.
They are a consequence of the phase transformation occuring on the microscale during cooling from
an initially stress-free austenitic material. To gain a more detailed understanding of the link between
production process and the final component’s properties, two computations are presented. First, the
cooling of a slice, representing a part of the cylindrical specimen is investigated with regard to evolving
macroscopic stresses. Additionally, an isothermal combination of Multi-Phase Field and Finite Element
simulations is used to analyze the influence of phase transformation regarding micro- and mesoscopic
residual stresses.
1 Cooling of a slice
A thermo-elasto-plastic Finite Element simulation is used to model the cooling process of the experimen-
tally investigated cylindrical specimen. This specimen has an eccentric hole to evoke a heterogeneous
distribution of the temperature during cooling. The cooling induces shrinking of the geometry, such that
the associated thermally induced strains are coupled to the mechanical strains, which are decomposed
into elastic and plastic strains. For the computation, an elasto-plastic material law with a von Mises
yield criterion and exponential hardening is used, following [5],
y(θ) = y0 (θ) + (y∞ (θ) − y0 (θ))(1 − exp(−δ)) + h . (1)
Therein, θ denotes the temperature, y0 the initial yield strength, y∞ the limiting yield strength, δ the
exponential hardening parameter, h the linear hardening parameter and the internal variable, which
resembles plastic effects. The boundary value problem bases on a slice of the cylindrical specimen.
∗ Corresponding author: sonja.uebing@uni-due.de
235
Time-temperature-relation at lateral surface

and temperature dependent volume expansion
0.1
z 1000
0.25 0.08
x
Temperature in °C
volume expansion
800
a) A−A
600 0.06
∂Bθ
400 0.04
17.5
y 200 0.02
8 ∂B A−A
x θ 0 0
in mm 0 10 11 12 13 14 15 20 30 40 50 60
b) 13 3.5 6 c) Time in s
Figure 1. a) and b): dimensions and boundary conditions for the slice using symmetry conditions. A−A illustrates
the investigated line. c) The slice is cooled over the lateral surface ∂Bθ ) from 1012.45 ◦ C down to room
temperature 20 ◦ C in 15 seconds and a volume expansion of 5% is applied.
Due to symmetry conditions, only one half of the slice is modeled, see Figure 1a,b, which illustrate the
dimensions and boundary conditions of the slice. Starting at an initial temperature of θ0 = 1012.45 ◦ C and
a pure austenitic material (austenitic volume fraction ϕ = 1), the lateral surface is cooled down to room
temperature of θ = 20 ◦ C, see Figure 1c. Between θ = 185 ◦ C and θ = 150 ◦ C the phase transformation is
considered by the material changing linearly to 87% martensite and 13% retained austenite (ϕ = 0.13).
Simultaneously, a volumetric jump of 5% is applied incrementally to represent the lattice shearing in a
simplified manner, see Figure 1c. On that account, the volumetric strain ε∗vol is applied on the principal
diagonal of the strain tensor, i.e.
0.05 1
ε∗vol = (185 ◦ C − θ), diag ε ⇐ diag ε − ε∗vol 1 . (2)
(185 ◦ C − 150 ◦ C) 3
For both materials, namely austenite and martensite, the temperature dependent material parameters
are calculated using Table 1 and the functions
f1 = ϕ(a1 + a2 θ)a3 + (1 − ϕ)(a4 + a5 θ)a6 , f2 = ϕ a1 exp(a2 θ)a3 + (1 − ϕ)a4 exp(a5 θ)a6 . (3)
Table 1. Material parameters w.r.t. temperature θ and austenitic volume fraction ϕ: bulk modulus κ, shear
modulus µ, initial yield strength y0 , limiting yield stress y∞ , exponential hardening δ, linear hardening h,
thermal conductivity aa, specific heat capacity cc, thermal expansion coefficient αT .
parameter unit f• a1 a2 a3 a4 a5 a6
κ MPa f1 1.01 −1.851 10−4 160287 1.015 −1.766 10−4 166349
µ MPa f1 1.024 −4.814 10−4 76700 1.025 −3.789 10−4 80977
y0 MPa f2 1.044 −1.475 10−3 420 1.028 −9.932 10−4 3050
y∞ MPa f2 1.044 −7.172 10−4 780 1.026 −7.437 10−4 3420
δ − f1 35 0 1 20 0 1
h MPa f1 520 0 1 500 0 1
aa N/(mm K) f1 0.966 6.868 10−4 1.758 10−2 1.016 8.024 10−5 3.721 10−2
cc J/(mm3 K) f1 0.986 3.862 10−4 3.692 10−3 0.946 9.550 10−4 3.548 10−3
αT 1/C f1 0.998 −9.180 10−5 2.361 10−5 0.986 2.889 10−4 1.228 10−5
When the temperature at the lateral surface, see Figure 1b, reaches 185 ◦ C (computation time t = 10 s),
the phase transformation starts. Up to now, no stresses occur on the investigated line, outlined in
Figure 1b as A−A with y = 0 mm and z = 0.125 mm, cf. Figure 2a. One second later, i.e. t = 11 s, high
compressive stress values arise near the boundary, cf. Figure 2b. The material core of the cylinder is still
warmer than the transformation start temperature of 185 ◦ C, such that no transformation has started.
At t = 15 s, the temperature at the lateral surface reaches room temperature. Now, the material core
transforms from austenite to martensite and the volumetric jump is applied. Thus, high compressive
stresses in this domain cause tensile stresses in the outer regions, in which the phase transformation is
already completed, see Figure 2c.
In comparison to Figure 1d, in which the results of the work at the IFUM, LU Hannover, are shown, it
is to be noticed, that the magnitude of the computed stresses highly differs. This has to be investigated
in detail such that the results from both approaches can be fitted accordingly. But qualitatively both
numerical investigations coincide such that tensile stresses occur at the outer boundary after the complete
cooling. Within the scope of the work to improve the component’s properties, e.g. compressive stresses in
the outer region, some adjustments to the process parameters should be done, e.g. modifying the cooling
route.
236
a) b)
350
Experiment
300
Simulation
250
200
150
100
50 216 338 138 157
thick thin
eccentric 0
hole MP 1 MP 2
c) d) Tangential stresses in MPa
Figure 2. Tangential stresses in MPa and temperature in ◦ C a) after 10 s, temperature at lateral surface θ = 185 ◦ C,
b) after 11 s, temperature at lateral surface θ = 150 ◦ C, c) after 15 s, temperature at lateral surface
θ = 20 ◦ C and d) tangential stresses in outer regions computed and measured by IFUM, LU Hannover.
2 Phase transformation on meso- and microscale
The numerical analysis of the slice only considers microscopic effects such as the phase transformation
in an approximate manner. Hence, the aim is to model the diffusionless phase transformation from
austenite to martensite occuring during cooling. Therefore, a combined model, firstly proposed in [1], of
an elastic Multi-Phase Field (MPF) model, see [6], and a two-scale Finite Element simulation, cf. e.g. [4],
is chosen. The MPF model is applied to describe the lattice shearing from face-centered cubic austentitic
unit cells to body-centered tetragonal martensitic unit cells. In order to describe the martensitic lattice
on the microscale three variants I, II and III are taken into account which are linked to eigenstrains
and represented by the so-called Bain groups, cf. [2], see Figure 3a. When applying a partial rank-
one homogenization scheme, martensitic grains containing only two of three variants are obtained, see
Figure 3b.
a) b) c) d)
Figure 3. a) Considered martensitic variants, b) microscopic laminate structure of one martensite grain, c) initial
and d) deformed mesoscopic polycrystal
To transfer the results to the mesoscale, another MPF simulation is used based on normal grain growth.
It configures an austenitic polycrystal, see Figure 3c, with a certain number of austenitic grains. The
polycrystalline structure is firstly deformed, see Figure 3d, and afterwards cooled down to room tem-
perature. During cooling the phase transformation from austenite to martensite occurs when reaching
the martensite start temperature defined as 185◦ C. The phase transfomation follows the principle of
minimizing the elastic energy in the system. Finally, depending on macroscopic, experimental results,
the polycrystal consists of approximately 87% martensite and 13% retained austenite.
With this information at hand, a two-scale Finite Element simulation [4] is done which takes into account
plastic effects and computes residual stresses of second type. More details are given in [1]. In absence of
outer forces, the resulting stresses can be interpreted as residual stresses rather than mechanical stresses.
The radial, tangential and axial stresses, denoted by σr , σϕ , σz , respectively as well as the equivalent
237
plastic strains εpeq and the von Mises stresses σvM are shown in Figure 4, of which the tangential stresses
are most important for the investigation of durability for instance.
σr σϕ
σz
a) b) c)
εpeq σvM
d) e)
Figure 4. Results: a) radial stresses σr in MPa, b) tangential stresses σϕ in MPa, c) axial stresses σz in MPa, d)
equivalent plastic strains εpeq and e) von Mises stresses σvM in MPa
In this work, the numerical analysis of a slice of cylindrical specimen, which was cooled from 1000 ◦ C to
20 ◦ C was presented. Therefore, a macroscopic and a micro-and mescoscopic model were applied in order
to compute residual stresses. The numerical simulations show qualitatively similar results comparing to
the experimental investigations. To obtain a more quantitative accordance, the coupling of the different
scales must be enhanced.
Acknowledgments
Thank you to: DFG - 374871564 (BE 1691/223-1, BR 5278/3-1, SCHR 570/33-1) for funding; Prof. Dr.-
Ing. B.-A. Behrens, Dipl.-Ing. A. Chugreev and C. Kock (IFUM, LU Hannover), Prof. Dr. I. Steinbach
and Dr. O. Shchyglo (ICAMS, RU Bochum) and Dr.-Ing. R. Niekamp and M. Sarhil (U Duisburg-Essen)
for their scientific collaboration and support.
References
[1] B.-A. Behrens, J. Schröder, D. Brands, L. Scheunemann, R. Niekamp, M. Sarhil, S. Uebing, and
C. Kock. Experimental and Numerical Investigations of the Development of Residual Stresses in
Thermo-Mechanically Processed Cr-Alloyed Steel 1.3505. Metals 9.4 (2019), 480 (28 pages).
[2] K. Bhattacharya. microstructure of martensite: why it forms and how it gives rise to the shape-
memory effect. Oxford: Oxford University Press, 2003.
[3] E. Macherauch, H. Wohlfahrt, and U. Wolfstied. Härterei-Technische Mittelungen - Zeitschrift für
Werkstoffe, Wärmebehandlung, Fertigung 28.3 (1973), pp. 201–211.
[4] J. Schröder. A numerical two-scale homogenization scheme: the FE2 -method. In: Plasticity and
Beyond - Microstructures, Crystal-Plasticity and Phase Transitions (CISM Lecture Notes 550, Eds.
J. Schröder, K. Hackl). Springer, 2014, pp. 1–64.
[5] J.C. Simo and C. Miehe. Associative coupled thermoplasticity at finite strains: Formulation, nu-
merical analysis and implementation. Computer Methods in Applied Mechanics and Engineering 98
(1992), pp. 41–104.
[6] I. Steinbach. Phase-field models in materials science. Modelling and Simulation in Materials Science
and Engineering 17 (2009), 073001 (31 pages).
238
MS 13: Advanced Numerical Methods for Fluid-Structure Interaction

A Weak Dirichlet-Neumann Partitioned Algorithm for

1-Dimensional Structures Embedded in a 3-Dimensional
Fluid
Nora Hagmeyer∗1 , Matthias Mayr1 and Alexander Popp1
1 Institute for Mathematics and Computer-Based Simulation, University of the Bundeswehr Munich, Germany
Abstract. Modeling of slender bodies via 1-dimensional beam theories can greatly increase the efficiency of
numerical simulations and lead to rather well-posed formulations. Especially the application of so-called ge-
ometrically exact beam theories as standalone models can yield precise solutions at only a fraction of the
computational cost of full 3-dimensional solid models. Harnessing the efficiency of these theories in mixed-
dimensional multi-physics simulations necessitates the development of efficient and stable load and motion
transfer schemes as well as constraint enforcement techniques for unavoidably non-matching meshes. This
contribution primarily addresses these challenges in the context of Fluid-Structure Interaction (FSI). In order
to couple a 1-dimensional structure model with the full 3-dimensional continuum theory of a surrounding
Newtonian incompressible fluid, a weak Dirichlet-Neumann coupling algorithm is proposed and applied to nu-
merical test problems. The focus lies on the transfer schemes and constraint enforcement techniques employed
to couple the 1-dimensional beam formulations and 3-dimensional fluid flow and their limitations.
Reduced Dimensional Structure Models
The interaction of rod-like structures with fluid flow plays an important role in a broad spectrum of
applications varying from biomechanical materials to industrial processes. In order to ensure a well-posed
formulation as well as an efficient usage of resources, applications can often gain from using 1-dimensional
beam theory to model the slender bodies as compared to a full 3-dimensional formulation.
In our case, the slender continua are modeled using geometrically nonlinear semi-induced beam theories
like the theories of Euler-Bernoulli and Kirchhoff, which consistently derive the kinematic relations from
the 3D continuum theory under some simplifying assumptions. On the other hand, the constitutive
equations are directly postulated in 1D and conserve essential mechanical principles such as equilibrium
of forces. For the numerical simulations using the Finite Element Method (FEM), the beams are dis-
cretized using a C 1 continuous centerline interpolation with Hermite polynomials, allowing for a smooth
approximation of the beam geometry even for coarse meshes [Meier2014].
(a) Depiction of the centerline in a beam with (b) Visualization of a Hermite beam centerline in-
circular cross section terpolation
Figure 1. Coarse interpolation of a slender body with 1D beam elements using 3rd order Hermite shape functions
for the centerline interpolation
∗ Corresponding author: nora.hagmeyer@unibw.de
239
Immersed Boundary Method
In terms of numerical modeling for the considered Fluid-Beam Interaction problem (FBI), the Finite
Element Method (FEM) applied to the incompressible isothermal Navier-Stokes equations as well as the
beam models is used in conjunction with an Immersed Boundary type approach. The Immersed Boundary
Method (IBM) was first introduced by Peskin in [Peskin2002] for the simulation of heart valves. These
types of methods have been the topic of ongoing research in recent years, in particular because of their
robustness in the presence of large displacements.
In Immersed Boundary type Methods, the boundary of the immersed solid is not resolved by the fluid
mesh, but a (volume) coupling of the fluid to the solid is implemented via a transfer of quantities, usually
load and kinematic values, between the two physics-specific meshes. The flexibility of this approach in
terms of the relation of the two meshes to each other makes it suitable for the application to mixed-
dimensional Fluid-Structure Interaction (FSI) problems.
A Weak Dirichlet-Neumann Partitioned Algorithm
In [Peskin2002], Peskin proposed a Neumann-Dirichlet type formulation in the sense that the fluid
velocity was transfered to the structure mesh in order to enforce the kinematic coupling, and vice versa
a force was transfered from the structure mesh and applied to the fluid. This builds a contrast to
Dirichlet-Neumann partitioned algorithms classically used in full-dimensional FSI simulation because of
their convergence properties. Reasons for this choice presumably include the problem of prescribing
velocity values within the incompressible fluid while still staying divergence-free in a weak sense. In
addition, when using simple linear material laws in the structure model, the fact that the fluid velocity
already completely determines the structure movement in each point makes it possible to condense the
structure equations, and to only supply the resulting interaction force to the fluid equation, thus avoiding
additional costly solution steps.
Nevertheless, we propose a Dirichlet-Neumann type partitioned algorithm that enforces the Dirich-
let conditions on the fluid in a weak sense. This allows to fulfill the divergence condition in the
vicinity of the embedded structure weakly and to treat occurring instabilities in a natural way us-
ing established coupling techniques. Similar Dirichlet-Neumann type algorithms have already been
proposed for Lagrange Multiplier based Mortar Finite Element surface coupling in contact mechanics
in [dirichletneumannconvergence] as well as in classical ALE-based Fluid-Structure Interaction in
[Kloppel2012]. Both works propose enforcing the Dirichlet condition, which is the non-penetration
condition in contact mechanics and the geometric continuity condition in FSI respectively, in a weak
sense, while also using the Lagrange multiplier to transfer the resulting interaction force to the Neumann
partition.
Constraint Enforcement Technique
There exists a multitude of techniques for the weak enforcement of Dirichlet conditions roughly separated
into the categories of Penalty Methods and Lagrange Multiplier based approaches, among them Nitsche’s
Method and stabilized Lagrange Multiplier approaches. Differing existentially in their derivation from
the weak from, members of these two categories display very different properties.
Within this work, a Penalty Method is employed to weakly enforce the kinematic coupling condition
on the Dirichlet partition, meaning that the following penalty term is added to the weak form of the
Navier-Stokes equations:
Z

v f − v̄ b · uf dΓ,
Γi
with the unknown fluid velocity v f , the projection of the beam velocity v̄ b in the fluid function space, the
fluid test function uf , the penalty parameter , and the FBI interface Γi . Besides the penalty method’s
simplicity, further advantages include its robustness, which comes with the price of never fulfilling the
enforced condition exactly for finite penalty parameters as well as the non-trivial choice of such a
problem-appropriate . Figure 2 depicts this fact for a rigid beam in a fluid channel.
240
Load and Motion Transfer Scheme
In order to obtain the projection of the beam velocity v̄ b , representing the weakly enforced Dirichlet
condition, into the fluid function space a projection operator has to be defined. In the discrete problem,
this corresponds to a mapping operator of the beam velocity to the fluid mesh. Within this contribu-
tion a Gauss-Point-to-Segment type approach, commonly employed in contact mechanics to transfer the
quantities of interest between the meshes, is applied to numerical examples showing the suitability of the
proposed method as well as its modeling limitations.
Figure 2. Fluid velocity in x-direction around a beam immersed in a fluid channel
241
242

Spline-Based Finite Element Method for Exact

Geometric Interface Representation
1
Michel Make ∗
, Norbert Hosters1 and Marek Behr1
1 Chair for Computational Analysis of Technical Systems (CATS), RWTH Aachen University, Germany
Abstract. Computer-Aided Design (CAD) tools are commonly used for the design of many engineering ap-
plications. These tools often utilize Non-Uniform Rational B-Splines (NURBS) to represent geometries. Iso-
geometric analysis (IGA) [3], allows us to directly use NURBS for numerical analysis. Analyzing fluid flows,
however, involves complex three-dimensional flow domains. Parametrizing such volumetric domains using
NURBS can be challenging and is still an ongoing topic in the IGA community.
With the recently developed NURBS-enhanced finite element method (NEFEM) [4, 5], the favorable geometric
characteristics of NURBS are utilized within a standard finite element method. This is achieved by enhancing
the elements touching the boundary by using the NURBS geometry itself. In the current work, the NEFEM
for a three-dimensional space-time formulation suitable for fluid-structure interaction simulations is discussed.
Furthermore, the method is demonstrated by means of a benchmark problems.
Introduction
In many engineering applications involving fluid-structure interaction (FSI) phenomena, accurate ge-
ometry representation is crucial. Computer-Aided Design (CAD) tools, due to their capabilities, are
commonly used in the early design stage of such applications. Such tools often rely on Non-Uniform
Rational B-Splines (NURBS) to represent geometries. The mathematical properties of NURBS make
them well suited to accurately describe complex shapes. The introduction of isogeometric analysis (IGA)
in [3], made it possible to directly exploit these favorable geometric properties for numerical analysis.
Analyzing fluid flows, however, often involves three-dimensional volume domains. Parametrizing such
domains using closed volume splines can be a challenge and is still an ongoing topic of research in the
IGA community.
To circumvent the need for volume splines, an alternative approach is proposed in [4], which is further
developed to incorporate space-time finite elements and free-surface flows in [5]. The method suggests the
use of standard finite elements in the interior of the computational domain, and supplement these with the
so-called NURBS-enhanced finite elements along the domain boundaries. These NURBS-enhanced finite
elements (NEFEM) use a NURBS description of the domain boundary in order to integrate the exact
geometry into the finite element method. By doing so, the error caused by discretizing a computational
domain by means of, e.g., an elementwise linear approximation is avoided (see Figure 1).
1 3D NURBS-Enhanced Finite Elements
Similar to its two-dimensional counterpart [2], for the 3D NURBS-enhanced elements, a non-linear map-
ping – termed Tetrahedron-Hexahedron-Tetrahedron (THT) mapping – from physical elements to the
reference element is used to incorporate the NURBS definition.
The THT-mapping is based on the idea to construct shape functions by degeneration [2]. First, a simple
mapping from a hexahedron to a tetrahedron is defined by coalescing nodes and adding together their
corresponding shape functions [6].
∗ Corresponding author: make@cats.rwth-aachen.de ORCID iD: 0000-0001-9504-955X
243
Figure 1. Standard FE and NEFEM mesh boundary.
To obtain the NEFEM, the mapping is first used to map a global tetrahedron element to a reference
hexahedron element. Using the inverse mapping, this hexahedron is then mapped to the well-known
reference tetrahedron (see Figure 2).
Figure 2. Tetrahedron-Hexahedron-Tetrahedron (THT) mapping.
The THT-mapping can now be used to define the shape functions and quadrature point distribution on
the elements which have a common face or edge with the NURBS boundary (see Figure 3). Note that
two different mappings need to be defined for the edge and face elements.
a) Element with NURBS face. b) Element with NURBS edge.
Figure 3. Different NEFEM element configurations.
The mapping for the face element shown in Figure 3a can be defined as follows:
!
~ 1 ξˆ + Θ
Θ ~ 3 η̂ + Θ
~ 4 ζ̂
ˆ η̂, ζ̂) = (ξˆ + η̂ + ζ̂)C
Φf ace (ξ, ~ + (1 − ξˆ − η̂ − ζ̂)~x2 . (1)
ˆ
ξ + η̂ + ζ̂
ˆ η̂ and ζ̂ represent the reference coordinates on the reference tetrahedron, Θ

Here, ξ, ~ i the parametric
244
coordinates of the element nodes along the NURBS and ~xi the global coordinates of the element nodes.
Note that the element node numbering in Equation (1) corresponds to that used in Figure 3.
Analogous to the method discussed in [2], the mappings used in this work can be extended for space-time
finite elements. This is done by defining a linear combination of of the upper and lower time level of a
space-time slab.
2 Domain Volume Computation
An accurate representation of the computational domain can significantly affect simulations involving FSI
phenomena, especially when closed domains are to be considered. This has been demonstrated previously
for two-dimensional problems in [1].
To show the benefit of the presented method, the volume of a domain with curved boundaries is studied.
In this case a cuboid is considered, from which a 120 degree segment is subtracted. The resulting volume
consists of five plane faces and one curved face, which can be represented by a NURBS surface exactly
(see also Figure 1).
Figure 4. Computational domain with curved face represented by a NURBS surface.
The volume of the domain is computed using both standard finite elements (SFEM) and NEFEM. For
NEFEM, the curved surface is represented by a second order NURBS surface with a control net as
shown in Figure 1. The absolute error between the exact volume and that computed using SFEM and
NEFEM elements is shown in Figure 5 and Table 1. The exact volume of the domain is defined by
Vexcact = L3 (2 − π9 − √148 ).
10−1 SFEM
NEFEM
Table 1. Grid refinement and relative error in volume
|Vnum −Vexact |
10−2
computation.
|Vexact |
10−3
Grid Elements Relative error
10−4
SFEM NEFEM
Rel. Err.
10−5 0 10 0.1140e+00 6.7198e−04

1 80 0.0336e+00 6.1552e−05
10−6
2 640 0.0087e+00 1.2086e−05
10−7 3 5,120 0.0022e+00 2.8218e−06
101 102 103 104 105 4 40,960 5.5479e−04 6.8374e−07
Number of Elements 5 327,680 1.3878e−04 1.6825e−07
Figure 5. Relative volume error for SFEM and NEFEM.
245
In Figure 5 and Table 1, it can be seen that the NEFEM formulation allows for an accurate volume
computation with a relatively coarse mesh. The computed error for NEFEM, relative to the exact
volume, is roughly four orders of magnitude smaller than that of the standard finite elements.
Acknowledgments
This work is partially made possible through the DAAD with funds from the Federal Ministry of Education
and Research (BMBF)
References
[1] N. Hosters. “Spline-based methods for fluid-structure interaction”. Dissertation. Aachen: RWTH
Aachen University, 2018. doi: 10.18154/RWTH-2018-223770.
[2] N. Hosters, J. Helmig, A. Stavrev, M. Behr, and S. Elgeti. Fluid–structure interaction with NURBS-
based coupling. Computer Methods in Applied Mechanics and Engineering 332 (2018), pp. 520 –539.
issn: 0045-7825. doi: https://doi.org/10.1016/j.cma.2018.01.003.
[3] T.J.R. Hughes, J.A. Cottrell, and Y. Bazilevs. Isogeometric analysis: CAD, finite elements, NURBS,
exact geometry and mesh refinement. en. Computer Methods in Applied Mechanics and Engineering
194.39-41 (Oct. 2005), pp. 4135–4195. issn: 00457825. doi: 10.1016/j.cma.2004.10.008.
[4] R. Sevilla, S. Fernández-Méndez, and A. Huerta. NURBS-Enhanced Finite Element Method (NE-
FEM). en. Archives of Computational Methods in Engineering 18.4 (Nov. 2011), p. 441. issn: 1134-
3060, 1886-1784. doi: 10.1007/s11831-011-9066-5.
[5] A Stavrev, P. Knechtges, S Elgeti, and A. Huerta. Space-time NURBS-enhanced finite elements for
free-surface flows in 2D. International Journal for Numerical Methods in Fluids 81.7 (2016), pp. 426–
450.
[6] O. C. Zienkiewicz, R. L. Taylor, and J. Z. Zhu. The finite element method: its basis and fundamentals.
Elsevier, 2005.
246
MS 14: Computational Contact & Interface Mechanics

An interface element formulation for frictionless normal

contact between rough surfaces
∗1
Jacopo Bonari , Maria Rosaria Marulli1 , Nora Hagmeyer2 , Matthias Mayr2 , Alexander Popp2 and Marco
Paggi1
1 IMT School for Advanced Studies Lucca
2 Institutefor Mathematics and Computer-Based Simulation, University of the Bundeswehr Munich
Abstract. A new multi-scale finite element formulation for contact mechanics between nominally smooth but
microscopically rough surfaces is proposed. The approach integrates the finite element method (FEM) for
modelling interface interactions at the macro-scale with a boundary element method for the solution of the
contact problem at the micro-scale.
Introduction
This contribution presents a novel finite element formulation for solving the problem of frictionless normal
contact between two bodies across a rough interface using a multi-scale approach. The method combines
a new interface finite element for modelling the surfaces interactions at the macro-scale with the boundary
element method (BEM) for the solution of the contact problem at the micro-scale.
The standard procedure for solving contact mechanics problem with roughness using the finite element
method (FEM) is limited by the issue of requiring an accurate representation of roughness. The proposed
variational approach overcomes this difficulty, treating the interface as nominally smooth at the macro-
level and embedding, at the micro-level, the actual rough topology, avoiding any assumption on the
surface height distribution, as previously done in other multi-scale frameworks [4].
1 Variational formulation
Let two elastic bodies occupy the domains Ωi ∈ R2 (i = 1, 2) in the not deformed configuration defined by
the reference system Oxz, Fig. 1. The two domains are separated
S by an interface Γ defined by the opposite
boundaries Γi (i = 1, 2) of the two bodies, viz. Γ = i=1,2 Γi , where contact takes place. The whole
boundary of the i-th body, ∂Ωi , is therefore divided into three parts: the classic Dirichlet and Neumann
boundaries, where displacements and tractions respectively are imposed, named as ∂ΩD i and ∂Ωi and
N
the interface Γi where specific boundary conditions have to be imposed in order to model contact. We
assume that Γi is nominally smooth but microscopically rough, resulting, in the not deformed state, in
two distinct but coincident curves. For each of the two bodies a displacement field ui = (ui , vi )T is
defined, while at the interface the configuration of the system is described by the gap field across the
interface g = (gt , gn )T , defined as the projection of the relative displacement u1 − u2 onto the normal and
tangential direction of the interface defined by the unit normal and tangent vector n and t. Neglecting
body forces, the strong form of equilibrium for each body reads:
∇ · σ i = 0 in Ωi , (1a)
ui = û on ∂ΩD
i , (1b)
σi · n = T N
on ∂Ωi , (1c)
gn ≥ 0, pn ≥ 0 on Γ. (1d)
∗ Corresponding author: jacopo.bonari@imtlucca.it
247
Γ1 Ω1
Ω2 Γ2
(e) S (e)
Γ = Γ1 S Γ2 Γ(e) = Γ1 Γ2
(e)
4 Γ1 3
n t
(e)
1 Γ2 2
Figure 1. Domains
S Ωi (i = 1, 2), their Dirichlet (∂Ωi ) and Neumann (∂Ωi ) boundaries, and the contact interface
D N
Γ = Γ1 Γ2 , together with the proposed interface discretization.
where û denotes the imposed displacement, and T the applied traction vector. The nominal pressure pn ,
is given by the micro-scale contact interactions predicted by the BEM, and can be qualitatively expressed
by: (
pn if gn > 0,
pn (gn ) = (2)
0, if gn ≤ 0.
2 Finite element discretization of the interface at the macro-scale
At the macro-scale, a four nodes zero thickness linear interface element is used for modelling surface
interactions. Its kinematics is borrowed from the formulation of interface elements used in non-linear
fracture mechanics for cohesive crack growth [2]. The integral expressions for the stiffness operator and
the residual vector are approximated via a Gaussian integration, and the values of the contact pressure are
evaluated by solving, for each Gauss point, the contact problem between an elastic half plane and a rough
surface, exploiting the BEM. To evaluate the normal gap gn at any point inside the interface element,
the nodal displacement vector d = (u1 , v1 , . . . , u4 , v4 )T must be introduced, collecting the displacements
u and v of the four nodes. The relative displacement g for the nodes 1-4 and 2-3 is given by:
g = −QNLd, (3)
where L maps the relative displacements, N interpolates them via linear shape functions over the length of
the element and Q rotates the reference system from the global to the local one. The integral expressions
for the stiffness operator and the residual vector are evaluated via a Gaussian quadrature and read:
2
X 2
X
R(k)
e = − det J wi LT NT QT p(xgi ), K(k)
e = det J wi LT NT QT C(xgi )QNL, (4)
i=1 i=1
√
where wi = 1 is the weight, xg1,2 = ∓1/ 3 are the positions of the two Gauss Points along the element,
p = (0, pn )T is the interface pressure and C is the linearised interface constitutive matrix:
 
0 0
C= ∂pn  (5)
0
∂gn
3 Boundary element method for micro-scale interactions
At this stage, the value of pn at each Gauss Point is computed by solving the normal contact problem of a
rigid rough surface indenting an elastic half-plane. The normal displacement at a point of the half-plane
248
ξ1 ξ2
gn
b
ū E, ν u
ξ2 ξ1
Figure 2. Illustration of the contact problem between a rigid rough surface, solid blue line, and an elastic half
plane, for a given far field displacement gn . The rigid body motion of the half plane is indicated by the
dashed black lines, while its deformed boundary by the solid black one.
characterized by a position vector ξ is related to the pressure p(η) exerted at other points by the following
relations: Z
u(ξ) = H(ξ, η)p(η) dη, (6)
S
where H(ξ, η) is the Green function, representing the displacement at point u(ξ) caused by a surface
pressure p(η). For a given far-field displacement gn , Eq. (6) must be satisfied, together with the unilateral
contact constraint [1]:
w(ξ, gn ) ≥ 0, (7a)
p(ξ) ≥ 0, (7b)
w(ξ, gn )p(ξ) = 0, (7c)
where w(ξ, gn ) = u(ξ) − u(ξ, gn ) and u(ξ, gn ) denotes the indentation of the half-space at the points in
contact, as shown in Fig. 2. The normal pressure pn is obtained dividing the total reaction force P by
the nominal surface area of the indenter.
4 Multi-scale coupling
Two main different approaches have been explored for performing the coupling between the micro- and
the macro-scale. In the first one, which is referred to as FEM-BEM semi-analytical (FBEM-SAN), the
normal contact problem at the micro-scale is solved off-line according to BEM, for a sequence of different
far-field displacements. The solution of the problem in terms of predicted average contact pressure vs.
the imposed far-field displacement is finally fitted with a power-law continuous function of the type:
pn (gn ) = agnb , (8)
which provides a closed-form expression for pn (gn ) [3]. Its derivative, ∂pn /∂gn , entering the linearised
interface stiffness matrix C, is also available in closed-form. In the second, a full integration of FEM and
BEM is proposed and labelled FEM BEM Quasi-Newton (FBEM-QN). The interface finite element has
been coded as a user element for the software FEAP, exploiting a Newton-Raphson solution scheme. At
each time step and for each Gauss point, the contact pressure pn (gn ) and the contact stiffness ∂pn /∂gn are
computed by calling twice the subroutine based on BEM: the first time pn is evaluated, the second one
a perturbation is applied to compute the normal contact stiffness via a finite difference approximation:
∂pn pn,k+1 − pn,k
' (9)
∂gn gn,k+1 − gn,k
where gn,k is the far-field displacement of the macro-scale model for the current k-th Newton-Raphson
iteration, and gn,k+1 = gn,k + ∆gn,k is a small perturbation of its value, for which the pressure values pn,k
and pn,k+1 are computed by BEM.
249
(a) Dimensionless overall reaction force against imposed (b) Point 3, A/An = 3.12%.
displacement.
Figure 3. Evolution of the free volume of the real geometry at the micro-scale.
5 Results
The results of a benchmark test involving two stacked blocks in contact, whose interface is discretized
by a single interface finite element are shown in Fig. 3 for a given imposed displacement ∆/s linearly
increasing between 0 and 3, being s the variance of the height distribution of the rough surface. The
scale separation is guaranteed by a ratio between the rigid indenter at the micro-scale and the interface
element size at the macro-scale of 0.1. In the contour plot, the ratio between the real contact area A and
the nominal one An is provided. The dark blue islands show the contact area while the deepest valley
still not in contact are marked in red.
Acknowledgments
The authors would like to acknowledge funding from the MIUR-DAAD Joint Mobility Program 2017 to
the project ”Multi-scale modelling of friction for large scale engineering problems”. The project has been
granted by the Italian Ministry of Education, University and Research (MIUR) and by the Deutscher
Akademischer Austausch Dienst (DAAD) through funds of the German Federal Ministry of Education
and research (BMBF).
References
[1] A. Bemporad and M. Paggi. Optimization algorithms for the solution of the frictionless normal
contact between rough surfaces. International Journal of Solids and Structures 69-70 (June 2015).
doi: 10.1016/j.ijsolstr.2015.06.005.
[2] P. Wriggers M. Paggi. A nonlocal cohesive zone model for finite thickness interfaces–Part II: FE
implementation and application to polycrystalline materials. Computational Materials Science 50.5
(2011), pp. 1634–1643.
[3] M. Paggi and J. R. Barber. Contact conductance of rough surfaces composed of modified RMD
patches. International Journal of Heat and Mass Transfer 54 (2011), pp. 4664–4672. doi: 10.1016/
j.ijheatmasstransfer.2011.06.011.
[4] G. Zavarise, P. Wriggers, E. Stein, and B.A. Schrefler. Real contact mechanisms and finite element
formulation–a coupled thermomechanical approach. International Journal for Numerical Methods in
Engineering 35 (1992), pp. 767–785.
250

Multiscale Rubber Friction Homogenization on Rough

Surfaces
Felix Hartung1 and Michael Kaliske ∗ 1
1 Institute for Structural Analysis, Technische Universität Dresden, 01062 Dresden
Abstract.
This paper introduces a multiscale approach for friction that considers friction on dierent length scale. On
each scale, a friction homogenization procedure generates the friction feature for the next coarser scale so that
a macroscopic friction coecient is accumulated out of the micro- and mesoscopic friction behavior. Within
the multiscale approach, adhesion implemented as non-linear traction-separation law is assumed to act only on
microscopic length scales. By using the nite element method (FEM), the sensitivity of the inuencing factors,
for instance slip velocity and contact pressure, may be investigated. A numerical validation of this approach
proves the validity of the homogenization method with respect to the distance of adjacent length scales.
Introduction
Friction, dened as force resisting the relative motion of solid surfaces, uid layers and material elements
sliding against each other, is a complex phenomenon and has a signicant importance in daily life. One
challenging eld of research is associated with tire industry, because tire-pavement interaction aects every
driving maneuver. Hence, the substrate structure is characterized by many asperities, it is important
to understand the physical background of sliding on dierent length scales. Friction in general consists
of dierent contributions - hysteresis friction, adhesion, viscous friction and interlocking eects. Both,
hysteresis friction due to internal dissipation of the viscoelastic material and adhesion in consequence of
molecular bonding to the surface represent the main friction parts. Adhesion as well as hysteresis friction
strongly depend on e.g. contact pressure, sliding velocity and temperature conditions. The analytical
approaches for multiscale friction of Persson (see among others [1]) and Klüppel/Heinrich (e.g. [2])
are common and show suitable results compared to friction measurements. Numerical models like the
multiscale approaches [3], [4] and [5], which base for instance on the nite element method (FEM), give
the opportunity to consider additional multi-physical phenomena.
1 Multiscale Approach
1.1 Scale Identication
Due to the dierent asperities of a rough surface, e.g. asphalt pavement, it is necessary to consider
the entire frequency spectrum or length scales. With several surface characterization methods, like the
height dierence correlation function (see [3] and [5]), the power spectral density function (see [4]) and
bandpass lters (see [6]), the frequency spectrum is classied in specic frequency clusters. In contrast
to the height dierence correlation and power spectral density function, bandpass lters provide the
decomposed surface textures for each length scale directly.
251
1.2 Adhesion Model
Adhesion describes mainly the molecular bonds between two dierent surfaces and gives an important
contribution to rubber friction. Adhesion behavior can be modeled for instance by a non-linear traction-
separation-law with the adhesional stress vector
σ = (1 − D) · K · δ , (1)
that is a function of the relative separation vector δ between two contact points in normal and two
tangential directions, the initial adhesional stiness K and the damage function
Z δmax
G 1
D= = σ̂ dδ . (2)
Gtot Gtot 0
In Eq. (2), G and Gtot represent the current and total fractal energy, δmax = max (kδk) is the largest
separation and σ̂ = kσk stands for the absolute adhesion stress. The adhesion model distinguishes
between normal (no contact) and tangential (σN = 0) adhesion. A patch test is shown in Fig. 1a in which
a cube with edge length of 1 mm is pressed by 1 N/mm2 on a rigid surface. Then the upper nodes (5-8) are
displaced 0.1 mm along the x-axis (step I). In the next step (II), the same nodes are moved back to the
original position. Finally, the upper nodes are lifted 0.1 mm along the z -axis (step III). The time duration
of each step is one second. The Coulomb friction coecient of the rigid surface is 0.5, K = 100 N/mm3
and Gtot = 0.01 N/mm.
(a) Patch test for adhesion model (b) Damage values and reaction forces during patch test
Figure 1. Patch test of adhesion model
Fig. 1b shows the damage evolution of all bottom nodes (1-4) and the resulting reaction forces. At the
beginning of step I, both tangential adhesion and friction are acting together until the total fracture
energy is reached. In step II, only friction remains, by changing the sign. At the beginning of step III,
friction vanishes, because of missing contact, and the adhesion model is switching from tangential to
normal adhesion. After the damage value is equal one, no stress in normal direction is transferred, the
contact points are released and the damage is reseted to zero to ensure repeated contacts.
1.3 Friction Homogenization
Within the multiscale approach, friction features as result of the deformation of dissipative material as
well as adhesion are homogenized on each scale to form a friction law for the next coarser scale. For
this purpose, FEM simulations are performed on block level. During every simulation, a rubber block
is pressed on a periodic rigid surface that is generated by the scale identication algorithm. Then, the
block is sliding over the rough rigid surface with a constant velocity. The ratio of the total horizontal and
252
vertical reaction forces forms the friction coecient µ (t) as a function of time. The time homogenization
algorithm of the friction coecient
Zttot
1
µhom = µ (t) dt (3)
ttot − tst
tst
only takes the steady state part starting at t = tst into account. In Fig. 2a, tst is reached after six periods.
The friction law for the next upper scale is computed by piecewise cubic spline interpolation
np nv
X X np −i nv −j
µspl (p, v) = ci,j · (p − ξp ) · (v − ξv ) (4)
i=1 j=1
with np = nv = 4, the breakpoints ξp and ξv (load and velocity conditions in each block simulation) and
the spline coecients ci,j . Fig. 2b shows the homogenized friction coecients at breakpoints ξp and ξv
as well as the spline evaluation (friction map). It has to be ensured that the friction map consists of an
(a) Homogenization of friction coecient (b) Friction map created by spline interpolation
Figure 2. Homogenization and friction map
adequate range of pressure and velocity breakpoints so that no friction coecients outside the tted map
are used during the block simulations on the next coarser length scale. On the microscopic length scale
(nest scale), only adhesion is applied as friction law. The computational costs strongly depend on the
scale identication algorithm, which inuences the mesh generation of the rubber blocks on each length
scale. Once the parameters K and Gtot are identied, the substrate is recorded in sucient detail and the
rubber material is characterized, this multiscale approach can compute macroscopic friction coecients
(adhesion plus hysteresis friction) as a function of macroscopic pressure and sliding velocity.
2 Numerical Validation of the Multiscale Approach
The multiscale approach is numerically validated by an simple 2D academic example using two dif-
ferent scales represented by sine waves (see also [4]). The macroscopic wave is described by z2 (x) =
0.075 mm sin (2 π/5 mm x) with a wave length of 5 mm. In contrast to [4], three dierent microscopic scales
z1,I , z1,II and z1,III with wave lengths of 0.5 mm, 0.2 mm and 0.1 mm and a ratio between amplitude
and wave length of 0.02 are investigated. Figs. 3a, 3b and 3c display the macroscopic friction coecients
of the full models with zi (x) = z1,i (x) + z2 (x) and i = I, II, III in comparison to the resulting friction
coecients of the macroscopic scale with and without microscopic friction. The macroscopic load and
slip velocity are 1 N/mm2 and 500 mm/s. The material properties as well as the mesh sizes are mainly
taken from [4]. Adhesion is neglected in this example. Even for this simple example, the multiscale
approach reduces the computational time by 86.76 % using z1,I , 97.28 % using z1,II and 99.50 % using
z1,III (assumption: all microscale simulation run in parallel) in comparison to the full model simulations.
The numerical validation example shows that the quality of the homogenization method depends on the
distance between adjacent scales. This should be considered within the scale identication to ensure the
validity of the multiscale approach.
253
(a) microscopic scale z1,I (b) microscopic scale z1,II (c) microscopic scale z1,III
Figure 3. Numerical validation of the multiscale approach using three dierent microscopic scales
3 Conclusion
The introduced multiscale approach can consider friction as result of deformation of dissipative materials
(hysteresis friction) as well as adhesion on dierent length scales. By homogenization of friction features on
each scale, a friction law as function of pressure and sliding velocity is obtained via piecewise cubic spline
interpolation for the corresponding coarser length scale. Note that the scale identication algorithm has
to ensure an adequate distance between adjacent length scales to guarantee the validity of the multiscale
homogenization procedure as explained in Section 2.
Acknowledgments
The nancial support of the German Research Foundation (DFG) of Research Unit FOR 2089 under
grant KA 1163/30 is gratefully acknowledged.
References
[1] B.N.J. Persson. Contact mechanics for randomly rough surfaces. Surface Science Reports 61 (2006),
pp. 201227. doi: 10.1016/j.surfrep.2006.04.001.
[2] A. Le Gal and M. Klüppel. Investigation and modelling of rubber stationary friction on rough
surfaces. Journal of Physics: Condensed Matter 20 (2007), p. 015007. doi: 10.1088/0953- 8984/
20/01/015007.
[3] P. Wriggers and J. Reinelt. Multi-scale approach for frictional contact of elastomers on rough rigid
surfaces. Computer Methods in Applied Mechanics and Engineering 198 (2009), pp. 19962008. doi:
10.1016/j.cma.2008.12.021.
[4] P. Wagner, P. Wriggers, C. Klapproth, C. Prange, and B. Wies. Multiscale FEM approach for hys-
teresis friction of rubber on rough surfaces. Computer Methods in Applied Mechanics and Engineering
296 (2015), pp. 150168. doi: 10.1016/j.cma.2015.08.003.
[5] K. Falk, R. Lang, and M. Kaliske. Multiscale Simulation to Determine Rubber Friction on Asphalt
Surfaces. Tire Science and Technology 44 (2016), pp. 226247. doi: 10.2346/tire.16.440401.
[6] P. Wagner, P. Wriggers, L. Veltmaat, H. Clasen, C. Prange, and B. Wies. Numerical multiscale
modelling and experimental validation of low speed rubber friction on rough road surfaces including
hysteretic and adhesive eects. Tribology International 111 (2017), pp. 243253. doi: 10.1016/j.
triboint.2017.03.015.
254

An unbiased self-contact formulation for explicit FEA

stabilized by the bipenalty method
∗1 1 1
Ján Kopa£ka , Du²an Gabriel and Radek Kolman
1 Institute of Thermomechanics of the Czech Academy of Sciences,Dolej²kova 1402/5, 182 00 Praha 8, Czech Republic
Abstract. In the explicit nite element analysis (FEA), contact boundary conditions are often enforced by
the penalty method. However, it is well known that the penalty parameter negatively aects the size of the
critical time step of the explicit time integration scheme. A remedy to this issue could provide the bipenalty
method. Recently, promising results for 1D contact-impact problems have conrmed this idea. Therefore,
further development and testing for higher spatial dimensions followed. The objective of this contribution
is to present the energy conservation properties of the bipenalty method and thus to prove the suitability
of this approach for solving the explicit FEA contact-impact problems. To this end, a symmetry preserving
contact algorithm has been modied to consider self-contact. Several numerical examples will be presented to
demonstrate the performance of the proposed contact algorithm.
Introduction
Computational analysis of fast contact-impact problems, such as drop tests of mobile or other devices,
crashworthiness of vehicles, or bird aircraft strike hazard is most often performed by the explicit nite
element method. Contact constraints are usually enforced by a contact-impact algorithm based on the
penalty method. The advantage of this approach is a simple implementation with the possibility of
parallelization. The disadvantage is then only the conditional stability of explicit time integration, which
dictates the maximum size of the time step. Moreover, the penalty method used to enforce the contact
constraints deteriorates the stability of the explicit time integration by increasing the maximum natural
frequency of the system. It is evident from the physical interpretation of the penalty parameter, which
can be understood as a spring that acts in the normal direction at the contact interface when penetration
occurs. The size of penetration is inversely proportional to the stiness of the penalty spring. The higher
the stiness of the spring, the smaller the penetration. Theoretically, zero penetration can be obtained
for innitely high stiness. In numerical practice, however, the innite value cannot be approached. The
high value of the penalty parameter leads to the poor conditionality of the linear system of equations for
implicit methods. For explicit calculation, it increases the maximum eigenfrequency of the system and
thus reduces the critical time step.
An alternative that eliminates the described diculties in explicit dynamics is the so-called bipenalty
method [1]. The main idea is to penalize not only potential energy but also kinetic energ. It leads not only
to the introduction of the stiness penalty but also the mass penalty parameter. By tuning their mutual
ratio, it is possible to inuence whether and how much the maximum natural frequency will increase.
Such a ratio of penalty parameters corresponds to the non-penalized system is called a critical penalty
ratio [4]. The bipenalty method was extended to systems with an arbitrary set of multipoint constraints
[5]. A detailed one-dimensional study of stability and reection-transmission properties of the bipenalty
method in contact-impact problems was performed in [6]. The bipenalty method thus appears to be a
valuable improvement to the standard penalties method.
In this paper, the bipenalty method will be combined with a symmetry-preserving contact formulation [3]
and used to solve a multidimensional impact problem. The potential of the proposed unbiased contact-
impact formulation to address self-contact tasks will be emphasized.
∗ Corresponding author: kopacka@it.cas.cz ORCID iD: 0000-0002-2975-8347
255
1 Contact formulation
The basis for the discretization of a problem described by partial dierential equations by the nite
element method is its formulation in a weak sense. The description of the contact formulation, therefore,
begins with a weak form of the balance of linear momentum which can be formulated as follows: in every
time instance, t, nd the displacement eld, u, from some suitably chosen function space, U , which for
all test functions, δv , from the selected test functions space, V , satises
2
(Z Z Z
X
δv (s) · ρü(s) dV + ∇δv (s) : P (s) dV − δv (s) · T c(s) dS
(s) (s) (s)
s=1 Ω0 Ω0 Γc
Z Z )
(s) (s)
− δv (s)
·T dS − δv (s)
·b dV =0 (1)
(s) (s)
ΓD ∂Ω0
where Ω0 is the domain representing the initial conguration, ρ is the density, u is the displacement
eld, P it the rst Piola-Kirchho stress tensor, T is the Piola traction vector, T c is the Piola contact
(s) (s)
traction vector, and b is the vector of volume forces. Further, Γc and ΓN denote contact and Neumann
boundary, respectively. Superimposed dots indicate time derivatives. The upper index, s, in round
brackets denotes the index of body.
In what follows, the attention will be paid only to the contact term which expresses the virtual power of
the contact forces
Z
δPcont := − δv · T c dS (2)
Γc
S (s)
where Γc = s Γc . One may notice that this term has the same structure as the Neumann boundary
condition. However, there is one dierence, neither the contact traction vector nor the contact boundary
is known in advance and as such are part of the solution. It leads to an optimization constrained problem.
The penalty method transforms this constrained optimization problem into a series of unconstrained ones
by penalizing the constraints, which at the end leads to the regularization of the contact traction vector.
In particular case of bipenalty method, the contact traction vector has the form
T c = (s hgN i + m H (gN ) g̈N ) n (3)

Notice that for the sake of simplicity of the presentation, only frictionless case is considered. Here, s
and m are the stiness and mass penalty parameters, respectively, n is the normal vector, H(•) is the
Heaviside step function, and h•i := (• + |•|) /2 are the Macaulay brackets representing the ramp function,
which return zero if the operand is negative and the operand itself otherwise. Finally, gN is the normal
gap function dened as gN := − (xm − x) · n, which for a point x returns the signed distance from the
closest point xm in the direction of n.
1.1 Spatial discretization
The spatial discretization is performed with the aid of the nite element method. Standard procedures
lead to the system of an ordinary dierential equation (ODE) in time
[M + Mcont ] d̈(t) + Finter (d(t)) + Fcont (d(t)) − Fextern (t) = 0 (4)
where M is the mass matrix, Mcont is the mass matrix due to contact inertia, d is the vector of equivalent
nodal displacements, and Finter , Fextern , Fcont are the internal, external, contact equivalent nodal forces,
respectively. Again, at this place, the attention will be paid only to the contact terms. They are assembled
from the local counterparts
Z
Fecont =− m hgN i GN dS (5)
Γec
256
Z

Mecont = m H (gN ) GN GT
Nm −GT
N dS (6)
Γec
where GN := Nn and GNm := Nm n. Here, matrices N and Nm consist of shape functions, Na ,

evaluated at the integration points and their closest point projections, xm , respectively. In particu-
T
lar N := N1 Insd · · · Nnnod Insd , where nnod is the number of nodal points of the element, nsd
indicates number of spatial dimensions, and Insd is the nsd -dimensional identity matrix.
1.2 Time discretization
The purpose of the time discretization is to transform the ODE system (4) into the system of algebraic
equations by means of an integration scheme. In this work, the central dierence method (CDM) is
employed. Its integration scheme is: Assume that the displacement vector dn at time step tn and half-
step velocity vector vn−1/2 are known. Then one can calculate the acceleration
−1
an = (M + Mcont ) {Finter (dn ) + Fcont (dn ) − Fextern (tn )} (7)
and update the velocity and displacement.
vn+1/2 = vn−1/2 + ∆tan , dn+1 = dn + ∆tvn+1/2 (8)
As already mentioned, this time integration scheme is conditionally stable. The stability analysis implies
the condition on the stable time step
2
∆t ≤ (9)
ωmax
where ωmax is the maximal angular eigenfrequency of the system.
2 Numerical example
The performance of the method will be presented by means of a numerical example of longitudinal impact
of two thick plates [2]. The geometry setting of the problem are depicted in Figure 1. The dimensions are
2d = 5 mm, and l = 2.5 mm. The linear elasticity is considered with material parameters: E = 2.1e5 MPa,
ν = 0.3, and ρ = 7800 kg/m3 . The inital velocity is v0 = 1 m/s and the stiness penalty parametr was
set to s = 1e16 MPa/m.
z y
v0 0
v0 x
2d
Figure 1. Longitudinal impact of two thick plates: geometry setting.
Figure 2 shows the dimensionless stress component, σp X = c1 σX /(Λv0 ), in the direction of the impact axis
∗
at dimensionless time t = c1 t/d = 0.56, where c1 = (Λ + 2µ)/ρ is the speed of the longitudinal wave,
∗
and Λ, µ are Lamé's constants. In the left part of Figure 2, one can see the convergence of the solution
with the mesh renement for the time integration with the critical time step. It can be seen that while
the standard penalty method would diverge, the bipenalty method is stable. In the right part of Figure 2,
one can see the convergence of the solution with the renement of the time step. One can also notice the
well-known fact that the oscillations caused by the spatial dispersion of the FEM mesh are minimal for
integration with the critical time step. Provided that the linear elements with the lumped mass matrix
are used.
257
0.5 0.5
0 0
−0.5 −0.5
−1 −1
σX
σX
∗
∗
−1.5 −1.5
−2 −2
−2.5 −2.5
−3 −3
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
∗ ∗
x x
analytic 50x50 analytic Cr = 0.125
25x25 100x100 Cr = 0.5 Cr = 0.99
Figure 2. The spatial (left) and temporal (right) convergence of the dimensionless stress component in the direction
of the impact axis.
3 Conclusions
In this paper, a symmetry-preserving contact formulation suitable for self-contact problems was presented.
Contact constraints were enforced by the bipenalty method, and the central dierence method was used
for time integration. A numerical example of the longitudinal impact of two thick plates was considered.
It was shown that by using the bipenalty method, it is possible to stably integrate even with a critical
time step corresponding to a non-penalized system.
Acknowledgments
The work of Radek Kolman and Ján Kopa£ka was supported by the grant project No. 19-04956S of the
Czech Science Foundation (CSF) within institutional support RVO:61388998 and the work of all three
authors was supported by the Centre of Excellence for Nonlinear Dynamic Behaviour of Advanced Mate-
rials in Engineering CZ.02.1.01/0.0/0.0/15_003/0000493 (Excellent Research Teams) in the framework
of Operational Programme Research, Development and Education.
References
[1] Harm Askes, Miguel Caramés-Saddler, and Antonio Rodríguez-Ferran. Bipenalty method for time
domain computational dynamics. Proceedings of the Royal Society A 466 (2010), pp. 13891408. doi:
10.1098/rspa.2009.0350.
[2] Rudolf Brepta and Franti²ek Vale². Longitudinal impact of bodies. Acta technica SAV 5 (1987),
pp. 575602.
[3] Du²an Gabriel, Ji°í Ple²ek, and Miran Ulbin. Symmetry preserving algorithm for large displacement
frictionless contact by the pre-discretization penalty method. International Journal for Numerical
Methods in Engineering 61.15 (2004), pp. 26152638. issn: 00295981. doi: 10.1002/nme.1173.
[4] Jack Hetherington, Antonio Rodríguez-Ferran, and Harm Askes. A new bipenalty formulation for
ensuring time step stability in time domain computational dynamics. International Journal for Nu-
merical Methods in Engineering 90 (2012), pp. 269286. doi: 10.1002/nme.3314.
[5] Jack Hetherington, Antonio Rodríguez-Ferran, and Harm Askes. The bipenalty method for arbi-
trary multipoint constraints. International Journal for Numerical Methods in Engineering 93 (2013),
pp. 465482. doi: 10.1002/nme.4389.
[6] Ján Kopa£ka, Anton Tkachuk, Du²an Gabriel, Radek Kolman, Manfred Bischo, and Ji°í Ple²ek.
On stability and reection-transmission analysis of the bipenalty method in contact-impact prob-
lems: A one-dimensional, homogeneous case study. International Journal for Numerical Methods in
Engineering 113.10 (2018), pp. 16071629. issn: 00295981. doi: 10.1002/nme.5712. url: http:
//doi.wiley.com/10.1002/nme.5712.
258
Pro eedings of the 8th GACM Colloquium on Computational Me hani s

for Young S ientists from A ademia and Industry
Extended Mortar Methods for Isogeometri Analysis

Melanie Krüger∗ 1 and Christian Hes h1
1 Chair of Computational Me hani s, University of Siegen
Abstra t. In the present ontribution, we rst introdu e a framework for large deformation porous-du tile
fra ture in thermo-elasto-plasti solids based on a modied Gurson-Tvergaard-Needelmann type plasti ity
model. Sin e higher-order phase-eld models require at least C 1 ontinuity, NURBS based shape fun tions
are applied for the spatial dis retization. Due to the tensor-produ t stru ture of the underlying spline bases,
severe restri tions in the geometry lead to a multi-pat h environment. In this ontext, we introdu e a novel
methodology, to enfor e interfa e onditions preserving higher-order ontinuity a ross the in general non-
onform interfa e, stated as extended Mortar method.
Introdu tion
The analysis of ra k initiation and propagation in du tile materials plays an important role in pre-
di ting failure me hanisms for various engineering appli ations. A large number of phenomenologi al
and mi rome hani al approa hes exists in literature for the modeling of du tile fra ture. Gurson devel-
oped a ma ros opi yield surfa e by homogenization of a porous RVE with assumed rigid plasti ow,
that degrades with in reasing void fra tion. This approa h has been further modied by Tvergaard and
Needlemann to a ount for damage growth, nu leation and oales en e ee ts. Here, we further extend
the oupled problem towards nite strain thermo-porous-plasti ity based on the phase-eld approa h.
Therefore, a multipli ative triple split of the deformation gradient, previously introdu ed in Dittmann et
al. [1℄, is presented along with a higher-order phase-eld formulation and a gradient.
A primal approa h in terms of onstrained basis fun tions ree ting weakly ontinuity up to maximal
regularity for multi-pat h geometries without rosspoints has been introdu ed in [2, 6℄ for problems in
stru tural me hani s su h as linear elasti ity and plate models. This formulation allows in a natural way
for a primal-dual reformulation in terms of a Lagrange multiplier. For this spe i ondition, we will
demonstrate dierent approa hes to enfor e the onstraints, either using a basis modi ation approa h,
a null-spa e redu tion s heme as well as the expli it usage of Lagrange multipliers. Afterwards, we
demonstrate how to deal with rosspoints in su h multi-pat h geometries, see [3℄. This modi ation
is onstru ted in su h a way, that we de ouple the Lagrange multipliers at the rosspoint to avoid a
global oupling ondition a ross all interfa es. Moreover, we re ast the underlying B-Splines su h that
they preserve the higher-order best approximation property a ross the interfa e and the rosspoint.
Eventually, we present a tual developments in the ontext of Mortar methodology applied to onta t
problems. Therefore, detailed investigation on thermome hani al porous du tile fra ture problems will
demonstrate the general appli ability, a ura y and apability of the proposed approa h, see [5℄.
In this se tion, we summarize the fundamental equations of the thermo-elasto-plasti solid in luding
porous du tile fra ture. The underlying, three-dimensional (d = 3) system leads to a multi-eld setting
with seven elds
{ϕ, s, θ, α, rp , F p , f } , (1)
∗ Corresponding author: melanie.kruegeruni-siegen.de
259
whi h are the deformation map, the ra k phase-eld, the absolute temperature and the four plasti
elds, whi h are the equivalent plasti strain, its dual hardening for e, the plasti deformation gradient
and the void volume fra tion. Here, the void volume fra tion is dened as
1 − f0
f =1− (2)
Jp
and is assumed as a mi ro-me hani ally motivated damage variable, related to the plasti deformation
gradient through its determinant J p . f0 is the initial void volume fra tion. The theory of Grith for
brittle fra ture [4℄ has to be modied to a ount for du tile behavior, i.e. the riti al lo al fra ture energy
density is dened as gc := gc (α). To implement the theory within a nite element framework, we dene
a regularized ra k surfa e fun tional, for the modeling of the ra k, whi h regularizes the sharp ra k
interfa e Γf0 with a spe i regularization prole γ
Z Z Z
gc (α) dΓ ≈ gc (α) γ(s, ∇s, ∆s) dV = Ψf (s, ∇s, ∆s, α) dV . (3)
Γf0 B0 B0
This regularization yields the ra k-density fun tion Ψf . The deformation gradient F = F e F p and its
determinant J = J e J p are de omposed multipli atively into elasti and plasti parts. Moreover, the
fra ture insensitive iso hori and volumetri parts are given by
(
Xh ig(s) (J e )g(s) if J e > 1
˜ e
F̄ = e −1/d e
(J ) λa na ⊗ N a and J˜ = e
, (4)
a
Je else
postulating that fra ture requires a lo al state of tensile/shear deformation. λea are the elasti prin ipal
stret hes, g(s) is a polynomial degradation fun tion and na and N a are the prin ipal dire tions of the
left and right stret h tensors.
The whole energy density with elasti , thermal, plasti and fra ture ontributions reads
Ψ := Ψe (F , F p , s, θ) + Ψθ (θ) + Ψp (α, ∇α, θ) + Ψf (s, ∇s, ∆s, α) , (5)
where the dierent dependen ies are indi ated.

Con erning the plasti material behavior, we postulate a modied Gurson model for porous plasti ity in
terms of the Kir hho stresses τ and the dissipative resistan e for e rp
2
σeq 3 p
Φp (τ , rp ) = p 2 + 2 q1 f cosh q2 p − 1 + [q1 f ]2 (6)
r 2 r
p
Here, the von Mises equivalent stress is represented by σeq = 3/2 kτ dev /Jk and the lo al pressure by
p = 31 tr[τ /J]. Moreover, q1 and q2 are tting parameters.
For the heat transfer, Duhamel's law of heat ondu tion is used and modied by the ra k phase-eld.
Here, the material thermal ondu tivity tensor is given by
K(F , s, θ) := {K0 [1 − wK (θ − θ0 )] (1 − s) + K conv s} C −1 . (7)
In ase of fra ture, we a hieve a pure onve tion problem and the heat transfer depends on the ra k
opening width. wK is a thermal softening parameter, K0 is a ondu tivity parameter related to the
referen e temperature, K conv is a onve tion parameter and C = F T F denotes the right Cau hy-Green
tensor.
2 Dis retization in spa e
Con erning the spatial dis retization, displa ement based nite element methods are applied subdividing
the domain B0 into a nite set of non-overlapping elements e ∈ E, su h that
[
B0 ≈ B0h = B0e . (8)
e∈E
260
Sin e the two elds of the plasti deformation gradient and the void volume fra tion are lo ally solved,
we have ve elds for the spatial approximation. These elds are approximated dierently. The approx-
imations of the solution eld of {ϕ, s, θ} and their variations read
X X
(•)h = RA (•)A , δ(•)h = RA δ(•)A , (9)
A∈I A∈I
Therein, we use the same spline based approximations for all three elds, where the global shape fun tions
RA : B0 → R are asso iated with ontrol points A ∈ I = {1, . . . , N}. With regard to the ontinuity
requirement of sh , δsh ∈ H2 (B0 ), we make use of quadrati NURBS based shape fun tions
Q
d
B ia (ξ a ) wi
a=1
A
R := R (ξ) = i
, (10)
P Qd
B ja (ξ a ) wj
j a=1
where B ia are univariate non-rational B-Splines and wj are orresponding NURBS weights. Note that
dierent shape fun tions are used for {α, rp } to avoid stability issues, see [1℄ for details.
3 Extended mortar method
Con erning a general multi-pat h situation with non- onformingly dis retized subdomains i ∈ {1, . . . , N },
the C n oupling onditions read
(i) (j) (i,j)
L(uh − uh ) = 0 on ∂Bh with i 6= j, (11)
(i,j) (i) (j)
where ∂Bh = ∂Bh ∩ ∂Bh denotes the ommon interfa e of two subdomains and the linear operator
L represents the respe tive oupling onditions† . Now, to enfor e these onditions we onstru t a saddle
point system by introdu ing a Lagrange multiplier eld on the ner resolved subdomain as
(i)
n̄
X
(i) (i) (i)
λh = RA,(i) (ξ̄ ) λA , (12)
A=1
(i)
where RA,(i) (ξ̄ ) are the B-spline based shape fun tions of the orresponding subdomain evaluated at
(i,j)
the interfa e ∂Bh . Here the number n̄(i) of involved Lagrange multipliers depends not only on the mesh
of the slave side but also on the order of the applied ontinuity. Applying a lassi al mortar approa h
for the implementation of C 0 ontinuity, the mortar onstraint for node A = 1, . . . , n̄(i) at the interfa e
(i,j)
∂Bh reads as
Z
A,(i,j) 2 (1−m) (i) (i) (i) (i) (j) (j)
Φ0 := hA RA,(i) (ξ̄ ) RB,(i) (ξ̄ ) uB − RA,(i) (ξ̄ ) RC,(j) (ξ̄ ) uC dΓ, (13)
(i,j)
∂Bh
A,(i,j)
whi h we enfor e su h that Φ0 = 0. Here, 2 m is the order of the onsidered dierential equation,
i.e., for a se ond order ellipti equation m = 1 the s aling fa tor is equal to one. The parametri B-spline
(j) (i)
oordinates ξ̄ are determined by a proje tion dened for a given ξ̄ as minimization problem
(i) (j) (i) (j)
kX̄ h (ξ (i) ) − X̄ h (ξ (j) )k = min kX̄ h (ξ (i) ) − X̄ h (ξ)k, (14)
(j)
ξ∈∂ B̃h
(i) (j)
where X̄ h and X̄ h represent the isoparametri geometry representation of the interfa e. It is obvious
that only basis fun tions having a non-zero tra e enter.
† E.g. Lu = u for a C 0 oupling, Lu = [u, u , u ]⊤ for a C 1 oupling and Lu = [u, u , u , u ⊤ for a C 2

,1 ,2 ,1 ,2 ,11 , u,22 , u,12 ]
oupling of two-dimensional subdomains. Here and in the following (·),i denotes the rst partial derivative with respe t the
ith-dire tion of the physi al spa e and (·),ij denotes the orresponding se ond partial derivative.
261
Next, for an extension of the mortar approa h to preserve C 1 ontinuity a ross the interfa e the mortar
onstraint at node A an be written as
A,(i,j) A,(i,j)
Φ1 := Φ0
d
X Z
2 (2−m) A,(i) (i) B,(i) (i) (i) A,(i) (i) C,(j) (j) (j) (15)
+ hA R,k (ξ̄ ) R,k (ξ̄ ) uB − R,k (ξ̄ ) R,k (ξ̄ ) uC dΓ
k=1 (i,j)
∂Bh
A,(i,j)
again enfor ed su h that Φ1 = 0. Here the s aling fa tor is equal to one for a fourth order equation
whereas the mesh size parameter hA has a negative power for sixth-order equations and a positive power
for se ond order equations. To guarantee C 2 - ontinuity weakly, we require for ea h node A
A,(i,j) A,(i,j) A,(i,j)
Φ2 := Φ0 + Φ1
d
X Z
2 (3−m) A,(i) (i) B,(i) (i) (i) A,(i) (i) C,(j) (j) (j)
+ hA R,kl (ξ̄ ) R,kl (ξ̄ ) uB − R,kl (ξ̄ ) R,kl (ξ̄ ) uC dΓ, (16)
k,l=1 (i,j)
k≥l ∂Bh
A,(i,j)
where we enfor e Φ2 = 0. For sixth-order equations the weak C 2 - ontinuity is mandatory. In theory
there is no limitation of this approa h and it an be naturally extended to C n - ontinuity, n > 2.
4 Conta t me hani s
Assuming two bodies in a onta t situation, the boundaries are no longer predetermined as usual for
domain de omposition problems. Moreover, the Lagrange multiplier eld represents the Piola tra tion
(i),c
t(i) asso iated with the onta t surfa e ∂B0 . In order to take dierent physi al me hanisms in normal
and tangential dire tion into a ount, the Piola tra tion is de omposed as follows
t(1) = −tN n − tT with tT · n, (17)
where n = n(2) denotes the urrent outward normal ve tor on ∂B (2),c . A ordingly, we dene the lo al
normal and tangential gap fun tions (gN , g T ), noting that we enfor e only C 0 ontinuity in normal
dire tion.
A knowledgments
Support for this resear h was provided by the German Resear h Foundation (DFG) under grant HE5943/8-
1 and DI 2306/1-1
Referen es
[1℄ M. Dittmann, F. Aldakheel, J. S hulte, P. Wriggers, and C. Hes h. Variational Phase-Field Formu-
lation of Non-Linear Du tile Fra ture. Comput. Methods Appl. Me h. Engrg. 342 (2018), pp. 71
94.
[2℄ M. Dittmann, S. S huÿ, B. Wohlmuth, and C. Hes h. Weak C n oupling for multi-pat h isogeometri
analysis in solid me hani s. Int. J. Numer. Methods Eng. (2019). doi: 10.1002/nme.6032.
[3℄ M. Dittmann, S. S huÿ, B. Wohlmuth, and C. Hes h. Crosspoint modi ation for multi-pat h iso-
geometri analysis. Comput. Methods Appl. Me h. Engrg. (submitted 2019).
[4℄ A.A. Grith. The phenomena of rupture and ow in solids. Philosophi al Transa tions of the Royal
So iety London 221 (1921), pp. 163198.
[5℄ M. Krüger, M. Dittmann, F. Aldakheel, A. Härtel, P. Wriggers, and C. Hes h. Porous-du tile fra -
ture in thermo-elasto-plasti solids with onta t appli ations. Computational Me hani s (submitted
2019).
[6℄ S. S huÿ, M. Dittmann, B. Wohlmuth, S. Klinkel, and C. Hes h. Multi-pat h isogeometri analysis
for Kir hho-Love shell elements. Comput. Methods Appl. Me h. Engrg. 349 (2019), pp. 91116.
262

Solving computational contact problems efficiently

at large scale
1
Matthias Mayr ∗
and Alexander Popp1
1 Institute for Mathematics and Computer-Based Simulation, University of the Bundeswehr Munich
Abstract. We present computational kernels and solution strategies for an efficient treatment of mortar terms
in the context of fully tied or unilateral contact mechanics problems. In particular, we propose approaches
to perform domain decompositions of mortar interface discretizations and discuss their implications on com-
putational cost and parallel scalability. We present strong and weak scaling results for mortar-related costs
in fully tied contact scenarios and exercise the interface domain decomposition capabilities in the context of
unilateral contact problems.
Introduction
Mortar finite element methods are used in a variety of interface problems, e.g. in fully tied or unilateral
contact, fluid flow, fluid-structure interaction, or electrochemistry [1, 2, 4, 5]. They allow for non-matching
interface meshes and enable a variationally consistent imposition of interface constraints. However, they
come at high numerical costs, which are in particular related to the numerical integration to form the
mortar projection operators and to incorporate mortar terms into the linear system. Also, the resulting
linear system requires specialized preconditioning techniques [6]. Here, we fully focus on the parallel
implementation and domain decomposition for an efficient evaluation of all mortar-related terms [3].
Problem description and finite element discretization
Two solid bodies Ω(1) and Ω(2) with unknown displacement fields u(i) , i ∈ {1, 2} and governed by the
laws of nonlinear elasticity interact with each other at the interface Γ∗ . At this interface, a kinematic
constraint g(u(1) , u(2) ) is imposed using a Lagrange multiplier field λ. Assuming suitable solution and
test spaces, the weak form of the coupled problem is given as
2
X Z Z
(i)
δW + λ δg dΓ + δλ g dΓ = 0.
i=1 Γ∗ Γ∗
Therein, the first term denotes the weak form of the PDE governing the domain Ω(i) , i ∈ {1, 2}, the
second term represents the interface traction contribution, while the third term expresses the kinematic
coupling constraint at the interface Γ∗ .
Using shape functions N to discretize the displacement fields u(i) , i ∈ {1, 2} and shape functions Φ to
discretize the Lagrange multiplier field λ, the mortar coupling matrices D and M are then assembled
∗ Corresponding author: matthias.mayr@unibw.de ORCID iD: 0000-0002-2780-1233
263
from the nodal contributions

D[j,k]
z" }| #{
Z m
XXn(1) (1) Z
∂g (1) (1)
λ δg dΓ ≈ λT
j Φj (1) u(1) , u(2) Nk dΓ δuk
Γ∗ j=1 k=1 Γ
(1)
∗,h
∂u
(1) (2) "Z # (1)
∂g (1) (2) (2)
m
XX n
T (2)
− λj Φj (2) u , u N` ◦ χh dΓ δu` .
j=1 `=1
(1)
Γ∗,h ∂u
| {z }
M[j,`]
Scalable evaluation of mortar matrices
Evaluating (1) is numerically expensive, especially since it requires projections of meshes and data between
the master and slave interface. To leverage the parallelism capabilities of modern computing hardware,
both domains Ω(i) , i ∈ {1, 2} are decomposed into subdomains Ω(m) , m ∈ {0, . . . , nproc − 1}. Then,
each subdomain is assigned to a parallel process m. To enable an efficient evaluation of the mortar
integrals (1), independent domain decompositions (DDs) are performed for the slave and master side of
the mortar interface, respectively. An exemplary DD for a two-dimensional problem run on four MPI
ranks is shown in Figure 1. Now, all processes m ∈ {0, . . . , nproc − 1} contribute their share to the
Figure 1. Example of independent volume and interface domain decompositions for optimal parallel processing.
Coloring reflects the assignment of a subdomain to a parallel process.
evaluation of (1), yielding speed-up of the entire mortar evaluation process. This process is often referred
to as load balancing.
Numerical results for the meshtying of two solid blocks
Strong and weak scaling studies for the time to evaluate (1) have been performed. For strong scaling, the
problem size is fixed, while the number of subdomains and parallel resources is increased, hoping for an
equivalent reduction in computational time. For weak scaling, the problem size is increased at the same
rate as the number of parallel processes, such that the work load per process is kept constant.
Figure 2 reports strong and weak scaling results for the meshtying of two 3D solid cubes. Optimal strong
and weak scaling can only be achieved, if the interface DD is independent from the underlying volume
DD.
Numerical results for the contact of a rolling cylinder
The cost of mortar integration is even more dominant in contact problems, since the mortar terms have
to be evaluated in every nonlinear iteration. Due to changes in the topology (either by rotation of a body
264
nel = 25 w/ LB nel = 100 w/o LB 12k w/ LB 28k w/o LB

el
n = 25 w/o LB nel = 125 w/ LB 12k w/o LB 60k w/ LB
nel = 100 w/ LB nel = 125 w/o LB 28k w/ LB 60k w/o LB
101 100
teval [s]
teval [s]
100
10−1
10−1
10−2
1 10 100 1 10
# processes [−] # processes [−]
(a) Strong scaling with nel elements along an edge (b) Weak scaling with 12k/28k/60k unknows per process
Figure 2. Scaling results for meshtying of two 3D solid cubes: strong and weak scaling can only be achieved, if
the interface DD (“w/ LB”) is independent from the underlying volume DD (“w/o LB”).
or by large sliding), the interface DDs depend on the current configuration. Thus, updating the interface
DDs is crucial for the overall efficiency of the contact solver.
We study three scenarios:
– Static LB: Perform an initial load-balancing as in the meshtying case.
– Static Contact-LB: Perform an initial load-balancing, but take contact search results into account.
– Dynamic Contact-LB: Monitor the load-balancing throughout the simulation and update the
interface DD if necessary while taking the contact search results into account.
Timing results for the mortar evaluation are reported in Figure 3. The work is only distributed equally
to all processes when using dynamic contact-specific load balancing. It also delivers the fastest mortar
evaluation throughout the entire simulation. A snapshot indicating volume and interface DDs for the
dynamic contact-specific load balancing is shown in Figure 4.
References
[1] A. Ehrl, A. Popp, V. Gravemeier, and W. A. Wall. A dual mortar approach for mesh tying within a
variational multiscale method for incompressible flow. Int. J. Numer. Meth. Fl. 76.1 (2014), pp. 1–27.
[2] R. Fang, P. Farah, A. Popp, and W. A. Wall. A monolithic, mortar-based interface coupling and
solution scheme for finite element simulations of lithium-ion cells. Int. J. Numer. Meth. Eng. 114.13
(2018), pp. 1411–1437.
[3] M. Mayr and A. Popp. “Scalable computational kernels for mortar coupling of interface problems”.
in preparation.
[4] M. Mayr, T. Klöppel, W. A. Wall, and M. W. Gee. A Temporal Consistent Monolithic Approach
to Fluid–Structure Interaction Enabling Single Field Predictors. SIAM J. Sci. Comput. 37.1 (2015),
B30–B59.
[5] A. Popp, M. Gitterle, M. W. Gee, and W. A. Wall. A dual mortar approach for 3D finite deformation
contact with consistent linearization. Int. J. Numer. Meth. Eng. 83.11 (2010), pp. 1428–1465.
[6] T. A. Wiesner, A. Popp, M. W. Gee, and W. A. Wall. Algebraic multigrid methods for dual mortar
finite element formulations in contact mechanics. Int. J. Numer. Meth. Eng. 114.4 (2018), pp. 399–
430.
265
Static LB Dynamic Contact-LB

Static Contact-LB limit: 1.8
teval max/min ratio [-]
12
10
8
6
4
2
0
0 50 100 150 200 250 300 350 400
time step [−]
(a) Parallel unbalance
Static LB Dynamic Contact-LB

Static Contact-LB
5
4
teval [sec]
3
2
1
0
0 50 100 150 200 250 300 350 400
time step [−]
(b) Contact evaluation time
Figure 3. Contact of a rolling cylinder: efficiency of various load rebalancing schemes.
Figure 4. Contact of a rolling cylinder: snapshot illustrating volume and interface DDs using dynamic contact-
specific load balancing.
266

Isogeometric Dual Mortar Methods for

Computational Contact and Interface Mechanics
1
Alexander Popp ∗
, Linus Wunderlich2 , Alexander Seitz3 and Barbara Wohlmuth4
1 Institute for Mathematics and Computer-Based Simulation, University of the Bundeswehr Munich, Germany
2 School of Mathematical Sciences, Queen Mary University of London, U.K.
3 Institute for Computational Mechanics, Technical University of Munich, Germany
4 Institute for Numerical Mathematics, Technical University of Munich, Germany
Abstract. Compared to classical Lagrange polynomials in finite element analysis (FEA), isogeometric analysis
(IGA) offers higher inter-element continuity and a smooth surface representation, which makes the applica-
tion to computational contact mechanics particularly appealing. In addition to contact analysis, so-called
weak patch-coupling is an important feature for practical applications of IGA. Within isogeometric methods,
the computational domain is usually divided into several spline patches and the solution is approximated
by spline functions on each patch. Within this contribution, a new class of mortar methods with dual La-
grange multiplier bases is transferred from the FEA to the IGA universe and for the first time combines
the efficiency of dual mortar methods with the beneficial IGA inherent concept of a smooth geometry. The
newly developed schemes are applied to both patch-coupling and unilateral contact problems. Spatial con-
vergence orders for the discretization error are studied numerically by applying uniform mesh refinement.
Isogeometric analysis
Robust and accurate contact discretizations for nonlinear finite element analysis have been an active
field of research in the past decade, and a new class of formulations has emerged with the introduction
of isogeometric analysis (IGA) [1]. IGA is intended to bridge the gap between computer aided design
(CAD) and finite element analysis (FEA) by using the smooth non-uniform rational B-splines (NURBS)
or T-splines common in CAD also as a basis for the numerical analysis. This promises some advantages
over classical Lagrange polynomials for FEA such as a possibly higher accuracy per degree of freedom
and, more importantly, a higher inter-element continuity. While finite elements based on Lagrange
polynomials are limited to C 0 inter-element continuity independent of the polynomial order p, NURBS
can be constructed with a maximum of C p−1 continuity. This high continuity results, amongst others,
in a smooth surface representation, which makes the application to computational contact mechanics
particularly appealing. In addition to contact analysis, so-called weak patch-coupling is an important
feature for practical applications of IGA. Within isogeometric methods, the computational domain is
usually divided into several spline patches and the solution is approximated by spline functions on each
patch. Typically, multivariate splines are defined based on a tensor-product structure, and a flexible
coupling between the individual patches is extremely important to gain some flexibility of the local
meshes (e.g. in order to meet different local mesh resolution requirements without graded meshes).
Mortar methods
Over the last two decades, mortar methods have emerged as one the most promising and mathematically
well-founded concepts to deal with non-conforming mesh- or patch-coupling and nonlinear frictional
contact problems in the context of classical FEA [3]. Consequently, there have been numerous efforts to
transfer the advantages of mortar methods to the IGA universe [2]. However, these efforts have so far
been limited to the class of so-called standard mortar methods, which simply construct a stable discrete
Lagrange multiplier basis using the trace space of the underlying mesh. While admittedly effective and
∗ Corresponding author: alexander.popp@unibw.de ORCID iD: 0000-0002-8820-466X
267
rather straightforward to transfer from the FEA to the IGA case, standard mortar methods cause high
computational costs due to the resulting indefinite equation systems of saddle point type. To propose
an alternative approximation of the Lagrange multipliers, so-called dual mortar methods were originally
introduced in the context of FEA domain decomposition [5] and later extended to finite deformation
contact analysis [3]. In contrast to standard mortar methods, the use of a dual Lagrange multiplier basis
yields a localization of the interface constraints due to the so-called biorthogonality property of the dual
shape functions. The aim of this work is to transfer this concept from finite elements to IGA, thus for
the first time combining the efficiency of dual mortar methods with the beneficial IGA inherent concept
of a smooth geometry. The newly developed isogeometric dual mortar method is applied to both patch-
coupling (i.e. domain decomposition) and contact problems. In both applications, spatial convergence
orders for the discretization error are studied numerically by applying uniform mesh refinement.
Figure 1. Primal (top), naive dual (middle) and optimal dual (bottom) basis functions shown on the first five
elements of a quadratic spline patch. The first and fourth basis functions are highlighted in bold.
A simple but non-optimal biorthogonal basis
As a first step, a rather straightforward and naive extension of dual mortar methods from FEA to
IGA is investigated: the dual basis functions in this first version have the same support as their primal
counterparts, fulfil a partition of unity and are constructed via element-wise linear combinations of the
primal basis functions [4]. Consequently, the global approximation order is limited to one in the L2 -norm,
independent of the local approximation. Therefore, since the dual NURBS do not possess the optimal
reproduction order, optimal convergence rates cannot be guaranteed theoretically. However, for contact
problems the solution is typically in H t (Ω(i))3 with t < 5/2, such that a priori estimates are already
limited by the regularity of the solution. It can be quite easily shown that the naive version of dual
mortar methods in IGA still yields optimal a priori estimates for the discrete contact problem. With
regard to NURBS patch-coupling, however, the naive version lacks the optimal approximation order such
that convergence of order O(hp ) in the H 1 -norm cannot be guaranteed, but only O(h3/2 ) can be achieved.
Extensive numerical tests and mesh refinement studies for both patch-coupling and finite deformation
contact within IGA have been carried out to confirm these theoretical findings in practice.
268
Figure 2. Exemplary convergence results for weak patch-coupling with quadratic NURBS: Comparison of the
naive biorthogonal basis (’ele dual’), optimal biorthogonal basis and standard Lagrange multipliers.
An optimal locally supported biorthogonal basis
To obtain optimal results with dual mortar methods in IGA also for the patch-coupling case, a new set of
biorthogonal splines is proposed as a second step [6]. The underlying idea comes from the fact that patch-
coupling optimality could be recovered for FEA by extending the support of the dual basis functions. An
extension of this approach to B-splines and especially its general implementation has still been an open
question until very recently. Here, we can finally present a new construction of local dual basis functions
with optimal approximation properties based on an extended support as suggested in the finite element
context. For the first time, this scheme allows to combine the two crucial features of local support of
the dual basis and optimal approximation properties. We first present the univariate construction (i.e.
1D interfaces), which has an inherent cross-point modification. The multivariate construction (i.e. 2D
interfaces) is then based on a tensor product approach for weighted integrals, whereby the important
properties are inherited from the univariate case. Numerical results include finite deformations in 3D
to confirm the optimal convergence properties in practice. They also demonstrate the suitability and
efficiency of the method for large-scale applications.
Figure 3. Exemplary contact problem with complex curved interface in 3D.
269
References
[1] J.A. Cottrell, A. Reali, Y. Bazilevs, and T.J.R. Hughes. Isogeometric analysis of structural vibrations.
Computer Methods in Applied Mechanics and Engineering 195.41–43 (2006), pp. 5257 –5296.
[2] L. De Lorenzis, P. Wriggers, and T.J.R. Hughes. Isogeometric contact: a review. GAMM-Mitteilungen
37.1 (2014), pp. 85–123.
[3] A. Popp and P. Wriggers, eds. Contact Modeling for Solids and Particles. Vol. 585. CISM Interna-
tional Centre for Mechanical Sciences. Springer International Publishing, 2018.
[4] A. Seitz, P. Farah, J. Kremheller, B. Wohlmuth, W.A. Wall, and A. Popp. Isogeometric dual mor-
tar methods for computational contact mechanics. Computer Methods in Applied Mechanics and
Engineering 301 (2016), pp. 259–280.
[5] B. Wohlmuth. A mortar finite element method using dual spaces for the Lagrange multiplier. SIAM
Journal on Numerical Analysis 38 (2000), pp. 989–1012.
[6] L. Wunderlich, A. Seitz, M.D. Alaydin, B. Wohlmuth, and A. Popp. Biorthogonal splines for opti-
mal weak patch-coupling in isogeometric analysis with applications to finite deformation elasticity.
Computer Methods in Applied Mechanics and Engineering 346 (2019), pp. 197–215.
270

A residual-type a posteriori estimator for the adaptive

discretization in space and time of a viscoelastic contact
problem.
∗ 1
Mirjam Walloth
1 Fachbereich Mathematik, TU Darmstadt, Germany
Abstract. We present a residual-type a posteriori estimator for the linear viscoelastic Signorini contact prob-
lem. The estimator constitutes global upper and lower bounds with respect to an error notion which measures
the error in the displacements, the velocities and a suitable approximation of the contact forces. For the spatial
error at each discrete time step local lower bounds are given in terms of the spatial estimator contributions.
As the estimator splits in temporal and spatial contributions it can be used for the adaptation of the time
step and the mesh size.
Introduction
As a contribution to the efficient numerical simulation of viscoelastic contact problems we present an a

posteriori error estimator.
We consider the linear viscoelastic Signorini contact problem as a model problem. The linear viscoelastic
material obeys the Kelvin-Voigt model. Viscoelasticity means that the material behavior shows both
elastic and viscous features and thus the stress tensor depends on the displacements as well as the
velocities. Examples of viscoelastic materials are biological tissue as cartilage, bones and tendons. The
resulting quasi-static contact problem gives rise to a time-dependent variational inequality.
Due to the constraints contact problems are non-smooth problems and thus the numerical simulation to
reach a certain accuracy is usually very expensive. Therefore, the adaptive numerical simulation based
on a posteriori estimators is in great demand.
A popular estimator for linear elliptic problems, which appears attractive in view of its simplicity and gen-
erality is standard residual estimation, see e.g. [5]. The estimator contributions are explicitly computable
from the given data and the discrete finite element solution.
In this talk we present a residual-type a posteriori error estimator for the time-dependent variational
inequality of the linear viscoelastic Signorini contact problem [6]. To the best of our knowledge that is
the first residual-type estimator for this kind of contact problem without a regularization of the non-
penetration condition. For a normal compliance viscoelastic contact problem we refer to [1].
We consider an error measure representing the error in the displacements, velocities and constraining
forces. The presented estimator constitutes a global in space and time upper bound and a global in
space and local in time lower bound of this error measure. Further, the spatial estimator contributions
constitute local lower bounds for the spatial error at each discrete time point. The details of the proofs
of efficiency and reliability are shown in [6].
The estimator splits in temporal and spatial contributions which can be used for the adaptation of the
time step as well as the mesh size. The estimator reduces to a standard residual estimator in the absence
of contact. The estimator contributions addressing the non-linearity are related to the contact stresses
and the complementarity condition with respect to the solution in two subsequent discrete time points.
Due to the fact that we exploit the local structure we avoid any spatial estimator contributions related
∗ Corresponding author: walloth@mathematik.tu-darmstadt.de
271
(a) deformation (b) mesh
Figure 1. Deformation and zoom in adaptively refined mesh
to the non-linearity in areas where the bodies are fully in contact in the surrounding of a node, so-called
full-contact areas. In consequence, the estimator perceives that at the boundary which is in full-contact
adaptive mesh refinement cannot improve the solution.
Finally, numerical examples confirm our theoretical results, see e.g. Figure 1. We present the convergence
rate of the estimator compared to uniform refinement. Further, the adaptively refined meshes at each
discrete time point and the adaptation of the time step size are shown.
1 Linear viscoelastic Signorini contact problem
We consider the contact between a linear viscoelastic body Ω and a rigid obstacle. The displacements
and velocities are denoted by u and u̇. The stress tensor obeys the following constitutive law σ :=
A(u̇) + B(u) where A is the viscosity tensor and B is the elasticity tensor and is the linearized
strain tensor. The time interval is denoted by I. Consistent to the linearized material, a linearized non-
penetration condition is considered. The linearized non-penetration condition is enforced by imposing
the constraints uν ≤ g for a given gap function g and a given direction of constraints ν on the potential
contact boundary ΓC . The Dirichlet and the Neumann boundary are denoted by ΓD , ΓN .
The complete problem formulation of the quasi-static contact problem is given by
Problem 1. Find a displacement field u : Ω × I¯ → Rd with u(0) = u0 such that
−divσ(u) = f in Ω×I
σ̂(u) = π on ΓN × I
u = uD on ΓD × I
uν ≤ g on ΓC × I
σ̂ν ≤ 0 on ΓC × I
(uν − g) · σ̂ν = 0 on ΓC × I
σ̂ tan = 0 on ΓC × I
with σ̂ := σ(u)n where n is the unit outward normal, σ̂ν := σ̂ · ν and σ̂ tan := σ̂(u) − σ̂ν ν.
1.1 Fully-discrete formulation
For the discretization in time we use the implicit Euler scheme and for the discretization in space we use
linear finite elements on a shape-regular mesh M with simplicial elements e. At each discrete time tn the
fully discrete solution is denoted by unm , the forces by f n = f (tn ), π n = π(tn ) and the gap function by
gmn
= gm (tn ). The space of linear finite elements in time step n is given by Hm := {v m ∈ C 0 (Ω) | ∀e ∈
272
M, v m,e ∈ P 1 (e) and v m = unD on ΓD } where we assume uD to be continuous and piecewise linear in
space and time. The admissible set is given by Knm := {v m ∈ Hm | vm,1 ≤ gm
n
on ΓC } where from now on
for the ease of presentation we choose the coordinate system such that ν = e1 .
For each time interval (tn−1 , tn ] we define the time step size τ n := tn − tn−1 and the linearly interpolated
n n n un −un−1
solution uτm := t τ −t n−1
n um + 1 − t τ −tn um and its derivative u̇τm := δunm := m τ n m . The fully discrete
problem is given by
Problem 2 (Discrete variational inequality). For each time step n = 1, . . . , N find unm ∈ Knm fulfilling
hA(δunm ), (v m ) − (unm )i + hB(unm ), (v m ) − (unm )i ≥ hf n , v m − unm i + hπ n , v m − unm iΓN , ∀v m ∈ Knm .
2 Residual-type a posteriori error estimator
The works [2, 4] reveal that sharp a posteriori estimators for obstacle and contact problems can be
derived by involving the error in the contact forces in the error measure. Here, the contact force density
in the continuous setting is given by hλ, ϕi−1,1 := hf , ϕi + hπ, ϕiΓN − hA(u̇), (ϕ)i − hB(u), (ϕ)i. To
compare the discrete contact force density hλnm , ϕm i−1,1 := hf n , ϕm i + hπ n , ϕm iΓN − hA(δunm ), (ϕm )i −
D τ E
hB(unm ), (ϕm )i to λ we need to extend it to a functional on H 1 . Therefore, we define λ̃m , ϕ =
D n E −1,1
λ̃m , ϕ on (tn−1 , tn ] with
−1,1
D n E X Z
λ̃m , ϕ := snp cp (ϕ1 ) φp
−1,1 bp,C
γ
p∈NsC
m
Z Z !
X
+ snp cp (ϕ1 ) φp − σ̂1 (unm )(ϕ1 − cp (ϕ1 ))φp (1)
bp,C
γ bp,C
γ
p∈NfC
m
hλnm,1 ,φp i−1,1

where cp (ϕ1 ) are weighted mean values, snp := R
φp
and γ
bp,C is the union of all boundary sides
γp,C
connected with p in time steps n and n − 1, compare [6, Section 2.4]. We distinguish between full-contact
n−1
nodes p ∈ NfCm , fulfilling um,1 = gm , um,1 = gm
n n n−1
and σ̂1 ≤ 0 on γ
bp,C and the remaining actual contact
nodes, called semi-contact nodes p ∈ Nm . sC
We note that compared to the work [3] on parabolic obstacle problems, the definition of full-contact
depending on the solutions at the discrete times tn and tn−1 enables to show lower bounds of η6 , η7 .
We define the global error measure on a time interval I which considers the error in the displacements,
the velocities as well as in the contact force density by
τ
ErrMeasG(uτm , u̇τm , λ̃m , I) :=
   D τ
E 2
Z hA(u̇ − u̇τm ), (ϕ)i + λ − λ̃m , ϕ
   −1,1 
 supϕ∈H0   (2)
I k∇ϕk
21
+k(u − uτm )k2L2 (I,L2 ) + k(u − uτm )(T )k2L2 ,
B A
where k · kL2A := hA(·), (·)i, k · kL2B := hB(·), (·)i are the norms induced by the bilinear forms.
Further, we define the local spatial error measure at each time tn
n
ErrMeasLp (unm , u̇τm , λ̃m ) :=
  D n
E 2  21
hA(u̇ − u̇τm ), (ϕ)i + λ − λ̃m , ϕ (3)
  −1,1  n 2 
supϕ∈H0 (ωp )   + k(u − um )kL2B (ωp )  .
k∇ϕkωp
273
The error estimator which we define in the following constitutes global upper and lower bounds with
respect to the global error measure (2) and local lower bounds with respect to the local spatial error
measure (3) at each time tn . The estimator
! 21 ! 12
X X
n 2
η := (η ) := (ητn )2 +τ n n 2
(ηm )
n n
q
τn
consists of the temporal estimator ητn := 3 k(um − um )kLB and the spatial estimator ηm :=
n n−1 2
n
qP
7
k=1 (ηk ) which has the following different contributions
n 2
! 12
P
η1n := n 2
(η1,p ) , n
η1,p := hp kf n kωbp
p∈Nm
! 12
P 1
η2n := n 2
(η2,p ) , n
η2,p := hp2 kJ I (unm )kγbp,I
p∈Nm
! 12
P
1

η3n := n 2
(η3,p ) , n
η3,p := hp2 J N (unm )
p∈NN̄ bp,N
γ
m
! 21
P 1
η4n := n 2
(η4,p ) , n
η4,p := hp2 kJ C n
tan (um )kγ
bp,C
p∈NC
m
! 12
P 1
η5n := n 2
(η5,p ) , n
η5,p := hp2 kσ̂1 (unm )kγbp,C
p∈NC fC
m \Nm
! 21
P R 12
η6n := n 2
(η6,p ) , n
η6,p := snp 12 bp (gm
n
− unm,1 )φp
p∈NsC
m
! 12
P R 12
n−1
η7n := n 2
(η7,p ) , n
η7,p := snp 12 bp (gm
n−1
− um,1 )φp
p∈NsC
m
where hp is the diameter of ω bp and bp ( ω bp . Here, the set of all nodes is denoted by Nm , the subset of
all nodes belonging to ΓC is given by Nm and the subset of all nodes belonging to Γ̄N is given by NN̄
C
m.
Further, J I (unm ) := (σ(unm )|e1 − σ(unm )|e2 ) · n are the jump terms where e1 , e2 are two elements sharing
a side and J N (unm ) := π − σ̂(unm ) and J C n n
tan (um ) := −σ̂ tan (um ).
If g 6= gm we get further estimator contributions, see [6, Section 3.2].
References
[1] J. R. Fernández and R. Martínez. A normal compliance contact problem in viscoelasticity: An a
posteriori error analysis and computational experiments. J. Comput. Appl. Math. 235.12 (2011),
pp. 3599–3614.
[2] Rolf Krause, Andreas Veeser, and Mirjam Walloth. An efficient and reliable residual-type a posteriori
error estimator for the Signorini problem. Numer. Math. 130.1 (2015), pp. 151–197.
[3] Kyoung-Sook Moon, Ricardo H. Nochetto, Tobias von Petersdorff, and Chen-song Zhang. A pos-
teriori error analysis for parabolic variational inequalities. M2AN Math. Model. Numer. Anal. 41.3
(2007), pp. 485–511.
[4] Andreas Veeser. Efficient and reliable a posteriori error estimators for elliptic obstacle problems.
SIAM J. Numer. Anal. 39.1 (2001), pp. 146–167.
[5] Rüdiger Verfürth. A posteriori error estimation techniques for finite element methods. Numerical
Mathematics and Scientific Computation. Oxford University Press, Oxford, 2013.
[6] Mirjam Walloth. Residual-type a posteriori error estimator for a quasi-static Signorini contact prob-
lem. IMA J. Numer. Anal. (2019). doi: 10.1093/imanum/drz023.
274
MS 15: Novel Formulations and Discretization Methods for Thin–walled Structures

Molecular Interactions of Slender Fibers Modeled as

Cosserat Continua
Maximilian J. Grill ∗ , Wolfgang A. Wall †
and Christoph Meier ‡
Institute for Computational Mechanics, Technical University of Munich, Boltzmannstr. 15, 85748 Garching b. München, Germany
Abstract. This contribution presents a 3D beam-to-beam interaction model for molecular interactions between
curved slender fibers undergoing large deformations recently proposed by the authors. A direct evaluation of
the total interaction potential for general 3D bodies requires the integration of contributions from molecule
or charge distributions over the volumes of the interaction partners, leading to a 6D integral (two nested
3D integrals) that has to be solved numerically. The novel strategy presented here proposes to formulate
reduced section-to-section interaction laws for the resultant interaction potential between a pair of cross-
sections of two slender fibers such that only two 1D integrals along the fibers’ length directions have to be solved
numerically. This section-to-section interaction potential (SSIP) approach yields a significant gain in efficiency,
which is essential to enable the simulation of relevant time and length scales for many practical applications.
1 Introduction
Motivated by the abundance and manifoldness of slender fibrous species (e.g. collagen, DNA, cellulose)
and their interactions in biological nano- and micro-scale systems, we propose a set of novel numerical
methods to model different kinds of molecular interactions between such fibers. Due to the focus on
macroscopic applications, most existing methods consider fiber interactions in form of mechanical contact,
interpreted as a non-penetration constraint (see e.g. [3, 6, 10]). On the scale of nano- and micrometers,
however, other molecular interactions such as electrostatics or van der Waals adhesion are no less crucial
for the system behavior and for instance even outweigh the effect of gravity by orders of magnitude.
Therefore, the development of accurate as well as efficient numerical methods is of utmost importance to
enable the simulation of the large and complex assemblies, e.g., networks, bundles or membranes, that
fibers typically form in biological systems. The long-term and higher-level goal of our research is to shed
light on some of the many yet poorly understood mechanisms and processes in the human body that are
governed by such fibrous structures. To give but one example, recent experiments with single collagen
fibrils suggest that molecular interactions between their constituting parts, which are again fiber-like
collagen molecules, contribute significantly to the overall mechanical response and might be exploited as
a mechanism to tune their structural stiffness [1]. The development of a computational model for collagen
fibrils in order to foster the in-depth understanding of their self-assembly and mechanics is one of the
biophysical applications we are currently pursuing in interdisciplinary collaborations.
2 The section-section interaction potential (SSIP) approach
Our approaches are based on the first principles of molecular interactions, which are formulated as point
pairwise interaction potentials Φ(r) and typically follow a power law in the separation r of the point-like,
fundamental interaction partners, e.g. molecules or point charges. Understanding two deformable 3D
bodies as continua composed of these fundamental interaction partners, the total interaction potential
∗ Corresponding author: grill@lnm.mw.tum.de ORCID iD: 0000-0003-1680-0259

† Corresponding author: wall@lnm.mw.tum.de ORCID iD: 0000-0001-7419-3384
‡ Corresponding author: meier@lnm.mw.tum.de ORCID iD: 0000-0002-3501-1696
275
requires the numerical evaluation of two nested 3D, i.e., a 6D integral over the contributions of all
individual point pairs [2, 8].
ZZ
Πia = ρ1 (x1 )ρ2 (x2 ) Φ(r) dV2 dV1 (1)
V1 ,V2
Unfortunately, these high-dimensional integrals come at a tremendous computational cost, which renders
the direct methods unfeasible if large fiber systems are considered. Thus, the slenderness of the fibers is
fully exploited by employing a mechanical description as 1D Cosserat continua widely known from the
so-called geometrically exact beam theory. In terms of the sought-after fiber-fiber interaction potential,
this results in our novel, so-called section-section interaction potential (SSIP) approach [4], where the
contributions of all pairs of beam cross-sections π̃ ˜ are cumulated in two nested 1D integrals along the
centerline curves r i of both interacting fibers i = {1, 2}.
ZZ ZZ
Πia = ρ1 (x1 )ρ2 (x2 ) Φ(r) dA2 dA1 ds2 ds1 . (2)
l1 ,l2 A1 ,A2
| {z }
˜ 1−2 ,ψ
SSIP law π̃(r 1−2 )
In this way, the need for numerical integration is reduced from 6D to 2D, which (in our considered
examples) saves a factor of 104 − 106 of integration point evaluations. The SSIP approach, which has
exemplarily been derived for the cases of Coulombic, van der Waals and repulsive steric interactions, thus
yields a significant gain in efficiency, which only enables the simulation of relevant time and length scales
in the applications described above.
3.1 Peeling and pull-off behavior of adhesive elastic fibers
The numerical example illustrated below has been studied in [5] and considers the separation of two
adhesive elastic fibers starting from contact along their entire length to fully separated fibers. Figure 1
shows the problem setup and resulting force-displacement curve.
y
ux , Fxtr in contact
normalized reaction force F̃x [-]
8
x separated
4
l
2
-2
ux , Fxbr 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
normalized displacement ux /l [-]
Figure 1. Numerical peeling and pull-off experiment with two adhesive elastic fibers. Problem setup (left) and
resulting quasi-static force-displacement curve (right). Force values to be interpreted as multiple of a
reference point load that causes a deflection of l/4 if applied at the fiber midpoint. Figure taken from [5].
3.2 Two oppositely charged fibers dynamically snap into contact
The numerical example illustrated below has been studied in [4] and considers two initially straight,
crossed fibers, which snap into contact due to the electrostatic attraction of their opposite surface charges.
Figure 2 shows the resulting sequence of simulation snapshots where electrostatic forces on both fibers
are shown in green.
276
Figure 2. Sequence of simulation snapshots (left to right, top to bottom). Adhesive electrostatic forces acting on
both fibers shown in green. Figure taken from [4].
4 Generalization
Interestingly, it will be shown that the section constitutive laws and objective deformation measures of
the geometrically exact beam theory according to Simo [9] and Reissner [7] can be considered as a special
case of the SSIP formulation in its most general form. In this special case, both interacting cross-sections
(and thus all molecular interaction partners) are parts of one and the same beam. In fact, this result is not
surprising since a body’s macroscale constitutive behavior eventually follows from microscale interactions
between its constituting molecules or atoms.
References
[1] Orestis G Andriotis, Sylvia Desissaire, and Philipp J Thurner. Collagen Fibrils: Nature’s Highly
Tunable Nonlinear Springs. ACS Nano 12.4 (2018), pp. 3671–3680.
[2] C Argento, A Jagota, and W C Carter. Surface formulation for molecular interactions of macroscopic
bodies. Journal of the Mechanics and Physics of Solids 45.7 (1997), pp. 1161–1183.
[3] D Durville. Contact-friction modeling within elastic beam assemblies: an application to knot tight-
ening. Computational Mechanics 49.6 (2012), pp. 687–707.
[4] Maximilian J. Grill, Wolfgang A. Wall, and Christoph Meier. A Computational Model for Molecular
Interactions Between Curved Slender Fibers Underging Large 3D Deformations With a Focus on
Electrostatic, van der Waals and Repulsive Steric Forces. In preparation ().
[5] Maximilian J. Grill, Christoph Meier, and Wolfgang A. Wall. Simulation of the Peeling and Pull-off
Behavior of Adhesive Elastic Fibers. In preparation ().
[6] Christoph Meier, Wolfgang A. Wall, and Alexander Popp. A unified approach for beam-to-beam
contact. Computer Methods in Applied Mechanics and Engineering 315 (2017), pp. 972–1010.
[7] E Reissner. On finite deformations of space-curved beams. Zeitschrift für Angewandte Mathematik
und Physik (ZAMP) 32.6 (1981), pp. 734–744.
[8] Roger A. Sauer and Laura De Lorenzis. A computational contact formulation based on surface
potentials. Computer Methods in Applied Mechanics and Engineering 253 (2013), pp. 369–395.
[9] J C Simo. A Finite Strain Beam Formulation. The Three-Dimensional Dynamic Problem. Part I.
Computer Methods in Applied Mechanics and Engineering 49 (1985), pp. 55–70.
[10] P. Wriggers and G Zavarise. On contact between three-dimensional beams undergoing large deflec-
tions. Communications in Numerical Methods in Engineering 13.6 (1997), pp. 429–438.
277
278

An isogeometric scaled boundary shell formulation for

thin-walled structures
1
Markus Klassen ∗
and Sven Klinkel1
1 RWTH Aachen University
Abstract. The present work aims at a numerical implementation of a shell formulation which is capable to
model extremely thin structures such as electroactive paper. To this end, the scale boundary concept is used
for the description of the shell kinematics. This concept leads to a separation of the shell deformation in an
in-plane and thickness direction. The separation is applicable to the discretization of the shell geometry as
well as to the solution method for both directions. By this means, the scaling parameter is associated with the
thickness of the shell. Furthermore, the formulation allows for a full 3D description of the constitutive law of
the shell material. In this contribution, the in-plane direction is treated numerically by isogeometric analysis.
This means that the in-plane geometry is discretized by NURBS functions which allow for a higher continuity
between the isogemetric elements. For the thickness direction of the shell, two alternatives are proposed. The
first option is given by a semi-analytical approach in which the displacement field along the thickness of the
shell is approximated by an analytical solution. Under the assumption of linear elastic material behavior, the
integration constants of the analytical ansatz are obtained by the boundary conditions of the top and bottom
surface of the shell. The second option is proposed by the use of a collocation method along the scaling
(thickness) direction. This approach can be regarded as a mixed numerical implementation with different
methods on the in-plane and thickness direction. Due to the geometric simplicity of the shell thickness, B-
Spline functions are considered for the numerical ansatz. The work presents the fundamental concept of the
scaled boundary method for shells. Details for an isogeometric discretization on the in-plane direction are
offered as well as the methodology for the two proposed shell thickness approaches. Subsequently, numerical
examples are provided in order to reveal the capabilities of the proposed methods. Therefore, the obtained
solutions are compared with analytical solutions as well as with standard finite element results.
1 Scaled boundary kinematics
∂3B A3 , ζ A3 , ζ
Ω
G1 G2 A2
A1
∂1B θ2
θ1
n̂1 n̂2
2
∂ B R
X
ei ei
Figure 1. Left: the shell body B determined by the boundary surfaces ∂ i B, right: the reference surface of the shell
with convective coordinates θα
The geometry of the shell body B is given by a curvilinear coordinate system θα , θ3 = ζ in which the
Greek indexes take the values of 1,2 while ζ is regarded as the thickness parameter which also corresponds
to the scaling parameter, see Fig. 1. A point on the reference surface Ω (mid surface) is defined by the
∗ Corresponding author: klassen@lbb.rwth-aachen.de
279
position vector R(θα ). Thus, by means of the scaled boundary geometry description, the position vector
of a point in the shell body B is provided according to,
X(θα , ζ) = R(θα ) + ζA3 (θα ), (1)
in which A3 is the normal vector to the reference surface Ω. The co-variant tangent vectors are obtained
by the parametric derivatives of the position vector. Thus,
Gα = R,α +ζA3 ,α , G3 = A3 in B. (2)
For the computation of the gradient and divergence of tensorial quantities in matrix notation by means
of the scaled boundary kinematics, the L-operator is introduced according to
 α   
G · e1 0 0 A3 · e 1 0 0
 0 G α
· e 0   0 A · e2 0 
 2   3 
 0 α  ∂. . .  0 A3 · e3 
L(. . . ) = 
0 G · e 3  +
0  ∂. . . . (3)
Gα · e2 Gα · e1 0  ∂θα  0 
 α  A3 · e2 A3 · e1  ∂ζ
G · e3 0 G α · e1  A3 · e3 0 A3 · e1 
α α
0 G · e3 G · e2 0 A3 · e 3 A3 · e2
| {z } | {z }
bα b3
It should be noted, that the matrices bα are defined by the contra-variant tangent vectors Gα and the
orthonormal Cartesian basis vectors ei . Under this perspective, the strain tensor is obtained in matrix
notation for the scaled boundary kinematics through the application of the L-operator on the displacement
vector u,
ε = Lu = bα u,α +b3 u,ζ . (4)
The equilibrium state of the shell body B is defined by the local balance of linear momentum and the
Neumann traction on the boundary of the shell ∂t B. Therefore, the equilibrium is modeled by the
governing field equations
LT σ + ρb = 0 in B, (5)
nTi σ = t̄ on ∂ti B. (6)
For the numerical solution of the boundary value problem, the displacement u and virtual displacement
δu are discretized for the in-plane directions by means of the introduced scaled boundary concept,
u(θα , ζ) = N (θα )û(ζ), δu(θα , ζ) = N (θα ) δ û(ζ). (7)
It should be noted, that the field vectors are approximated by form functions N (θ ) which are defined α
on the reference surface Ω of the shell. The discretization in scaling (thickness) direction û(ζ), δ û(ζ) is
not defined yet, since two approaches are to be compared in this contribution. For this reason, the field
equation in scaling direction needs to derived by taking the discretization on the in-plane direction into
account. This is performed through the application of the principle of virtual work on eqns. (5) and
(6) and subsequent partial integration. Furthermore, a linear elastic constitutive law is assumed in the
present contribution. Substituting the discretizations (7) in the derivation process leads to the balance
equations for the scaling direction which are obtained as,
T
E 0 û,ζζ +(E 1 − E T1 − Ē 1 )û,ζ −(E 2 + Ē 2 )û + F b = 0, (8)

(E 0 û,ζ +E 1 û − F ζ ) = 0. (9)
3
∂ B
It is observed, that the differential equations are functions of the scaling parameter ζ. The coefficients
are obtained by the integration over the reference surface Ω and are defined as follows:
∫ ∫ ∫ ∫ ∫ ∫
E0 = T
C 1
B 3 B 3 J dθ dθ ,2
E1 = T 1 2
C
B 3 B 12 J dθ dθ , Ē 1 = B T3 B̄ 12 J dθ1 dθ2 , (10) C
θ1 θ2 θ1 θ2 θ1 θ2
∫ ∫ ∫ ∫
E2 = C
B T12 B 12 J dθ1 dθ2 , Ē 2 =
T
C
B̄ 12 B 12 J dθ1 dθ2 . (11)
θ1 θ2 θ1 θ2
280
The C matrix denotes the elasticity tensor for a 3D linear elastic constitutive law and J stands for the
determinant of the Jacobian. The B-matrices are obtained in dependency of the in-plane shape functions
N (θα ) and the matrices introduced with the L-operator,
T
B T3 = N T bT3 , B T12 = N T ,α bTα , B̄ 12 = N T bTα ,α . (12)
The shell reference surface Ω is provided by CAD data. By means of isogeometric analysis, the dis-
cretization of the solution field on the in-plane direction, see (7), is performed with NURBS functions.
Regarding the scaling direction, two approaches are considered.
3 Methodology in scaling direction
In order to complete the element formulation, two methods are proposed for the solution of the thickness
direction of the shell. The first method consists in a semi-analytical formulation in which the thickness
displacement distribution is described by an exponential function. The second approach is a numerical
one which is defined by the collocation method which solves the governing equations at the collocation
points.
3.1 Semi-analytical method
A semi-analytical shell element formulation is obtained by providing the analytical solution for eqns. (8)
and (9). To this end, the mentioned equations are rearranged into a system of first order differential equa-
tions. In addition, a dual vector field is introduced as the main unknown. It consists of the displacement
field û(ζ) and the traction force F ζ (ζ) in scaling direction. Consequently, the system of equations can be
expressed as a homogeneous first order differential equation in which the matrix exponential function is
the general solution. The matrix exponential function is computed by a Taylor series. The consideration
of the traction boundary conditions F ζ (ζ) at the top and bottom surface of the shell body allows for the
elimination of the integration constants in the general solution. Thus, the relation between F ζ (ζ) and
û(ζ) is obtained, which defines the shell element stiffness matrix, see [1]. Each element node is defined
by 6 degrees of freedom in which three correspond to the displacement of the shell bottom and three to
the shell top.
3.2 Collocation Method
In the collocation method, the field eqns. (8) and (9) are evaluated at so called collocation points along
the scaling direction. For this purpose, the displacement field û(ζ) is discretized by a linear combination
of B-Spline functions which is then substituted in the mentioned equations. The order of the B-Spline
functions is defined by the number of collocation points according to p = ncp − 1 in which ncp is the
number of collocation points. The coordinates of the collocation points are defined in dependency of the
B-Spline knot vector by means of the Greenville abscissae. By the evaluation of eqns. (8) and (9), a
linear system of equations is obtained which defines the stiffness matrix of the shell element. In contrast
to the semi-analytical shell element, the number of degrees of freedom per node depends on the number
of collocations points in which three degrees of freedom are assigned to each collocation point.
4 Numerical results
The objective of the numerical example is to provide a first implementation test of the proposed iso-
geometric shell formulation with the analytical and collocation method for the thickness direction. To
this end, the proposed methods are compared with the implementation of a standard solid finite element
and a scaled boundary method in which the thickness direction is solved by a weak formulation. For
the sake of simplicity, a plate geometry as depicted in Fig. 2 is studied. The dimensions of the plate are
defined by w = l = 1 and the thickness is considered as a variable parameter for the study. A linear
elastic constitutive law is defined for the material behavior with E = 1000 and ν = 0.3. Furthermore, a
traction of Tx = 4 is applied as a boundary condition as shown in Fig. 2. It is important to note, that
the traction boundary condition is only applied to the degrees of freedom which correspond to the top
281
Figure 2. Setting. Figure 3. Results
and bottom surface of the plate. The reason is given by the limitation of the semi-analytical element
which only allows for top and bottom degrees of freedom. In order to evaluate the numerical results, a
relative error is defined for the study. It consists of the relative difference between the strain in traction
direction of the numerical results and the strain obtained according to Hooke’s law for a linear elastic
material. At this point it should be noted, that the analytical strain obtained by Hook’s law assumes a
homogeneous traction distribution over the thickness of the plate. Since the traction forces are applied as
singular loads on the top and bottom surface of the plate in the proposed methods, an inward buckling
of the cross section is predicted. This buckling effect is more pronounced for thick plates and vanishes
with decreasing plate thickness.
The discretization of the plate is performed on the in-plane direction with linear NURBS which corre-
sponds to a standard finite element implementation with linear Lagrange polynomials. In scaling direc-
tion, four different strategies are considered. The first one considers the semi-analytical (SA) method
while the second one consists in the collocation method with 3 (CLC3) and 5 (CLC5) collocation points.
In both strategies, the traction forces are only applied on the top and bottom degrees of freedom which
allows for the cross sectional buckling. In the third case, solid elements (Solid) with linear Lagrange
interpolation are considered while the fourth method is defined by a weak formulation with two and
three nodes over the plate thickness with linear (WF2) and quadratic (WF3) interpolation. In the latter
strategies, the application of the traction force is applied by means of a constant distribution over the
thickness which does not allow for a cross sectional buckling.
The relative error in dependency of the length to thickness ratio is depicted in Fig. 3. It can be observed,
that the semi-analytical (SA) and the collocation method (CLC3, CLC5) converge to the analytical so-
lution with a decreasing plate thickness. This effect is due to the traction boundary conditions which
allow for a cross sectional buckling. For the cases of a large plate thickness, the buckling leads to a larger
deformation than the one predicted by Hooke’s law for constant traction over the thickness. Furthermore,
it is observed that the proposed methods agree very well. In the simulations with a constant traction
over the thickness (Solid, WF2 and WF3), the relative error increases unsteadily as the plate thickness
decreases. This observation speaks for a lower capability of the latter methods to handle elements with
a large length to thickness ratio.
Acknowledgments
The authors thank the German research foundation (DFG) for the funding of the project KL 1345/15-1.
References
[1] H. Man, C. Song, W. Gao, and F. Tin-Loi. A unified 3D-based technique for plate bending analysis
using scaled boundary finite element method. Int. J. Numer. Meth. Engng 91 (2012), pp. 491–515.
282

Free Material Optimization of Multilayer

Composite Materials
Simon Loske ∗1
1 Structural Mechanics, TU Dortmund, Dortmund, Germany
Abstract. The use of composite materials in various types of devices has increased steadily in the past decades.
Hence, it is worth investigating into the structural behaviour of these materials. The optimization of stacking
sequences in multilayer composite materials opens a large field of improvements in industrial applications of
composite materials. The center of investigation is set on fiber reinforced laminate plies, which are examined in
the case of linear anisotropic theory of elasticity. One very auspicious way for optimization of these laminates is
the implementation of Free Material Optimization (FMO). In these optimization procedures, arbitrary values
of parameters of the elasticity matrix are obtained under constraints that only require the elasticity matrix
to be symmetric and positive semidefinite. Also, the trace of the elasticity matrix is restricted to prevent the
total stiffness of the material from becoming unnaturally large. Sensitivity information is obtained by the
variational approach for sensitivity analysis [1], which represents the most efficient known way of computing
sensitivity information.
Introduction
The most important application area of FMO is thin walled structures like laminates of very high strength
fiber materials, often used in aerospace or wind energy applications. The fundamental idea of FMO is
the optimization of a material tensor, which can vary in each single Gaussian point of the design domain.
There is only one major restriction to the optimization, which is that the material tensor has to be positive
semidefinite. The obtained values of the material tensor are characterizing the distribution of material
density and local material properties in an indirect way. The most promising approach with the idea of
FMO lies in patch field analysis, which identifies larger areas in the structure with insignificantly changing
stress fields. Due to the restriction on the material tensor to be positive semidefinite, the obtained
problem is a nonlinear semidefinite optimization problem [3]. With that requirement, the Method of
Moving Asymptotes (MMA) [2] seems to be an appropriate alternative to the nonlinear semidefinite
optimization algorithms. Controlling the eigenvalues of the material tensors to be strictly positive values
fulfills the controlling of positive semidefiniteness.
1 FMO Problem Formulation
The examined optimization problem is limited to two-dimensional linear elasticity, so that stresses can
only occur in the plane. Assuming a two-dimensional material law reduces the number of independent
material parameters in each Gaussian point from a total number of 21 (symmetric three-dimensional
material) to a number of only 6 independent parameters
 
E1 E3 E5
T
Ê = E3 E2 E6  , with Ê = Ê < 0. (1)
E5 E6 E4
In the FMO approach the material parameters are usually not restricted by physical properties. It is
convenient to restrict the design domain by boxed constraints that are calibrated by reasonable values
which are known from the typical physical behavior of materials like Young’s modulus or Poisson’s ratio.
∗ Corresponding author: simon.loske@tu-dortmund.de
283
However, there are no restrictions to the relation between different parameters, which are part of the
material tensor, so that they can vary completely independently from each other. That leads to the fact
that, normally, no material exists that can exactly fulfill all the optimized material parameters. This
is not a big problem if the focus is put on multilayer materials. In this case, the material tensor of a
laminate is the result of preintegration of material properties in each single layer over the whole thickness.
The weak form of static equilibrium can be obtained by
Z Z Z
R(u, v) = σ̂(u, s) : ε̂(v) dΩ − b · v dΩ − t̄ · v dΓ. (2)
Ω Ω Γ
With this weak form, the pseudoload matrix, which is essential for the optimization can be calculated by
variation. In this weak form only the term of stresses is dependent on the design variables
Z Z
∂R ∂ σ̂(u, s) ∂ Ê
P = = : ε̂(v) dΩ = ε̂(u) : ε̂(v) dΩ. (3)
∂s Ω ∂s Ω ∂s
The optimization problem is the minimization of compliance,that is, the maximization of global stiffness.
The compliance C and its gradient with respect to changes in design can be calculated from known
quantities
∂C(s) ∂Π(s)
C(s) = −2 Π(s) = uT f , ∇s C(s) = = −2 = −uT P . (4)
∂s ∂s
This leads to the optimization problem with control of eigenvalues and density
minimize C(s),
sli ≤ si ≤ sui , i = 1, ..., 6,
−λj ≤ 0, j = 1, ..., 3, (5)
3
X
ρl ≤ t λ j ≤ ρu .
j=1
2 Numerical Example With Patch Analysis
To show that the optimization scheme is working precisely, a numerical example is consulted. Therefore,
a cantilever beam is chosen, which is clamped on one edge and loaded on the other. For the initial design
and the affecting force, the following values are assumed
0.35
2.5 0.35
2 0.3
0.3
2
1.5 0.25
0.25
1.5
0.2
1 0.2
1
0.15
0.5 0.15
0.5
0.1
0.1
0
0
0.05
0.05
-0.5
0 -0.5
0
0 0.5 1 1.5 2 2.5 3 3.5 4
0 0.5 1 1.5 2 2.5 3 3.5 4
Figure 1. Initial deformation Figure 2. Optimized deformation

of structure of structure
284
 
50000 1000 1000
Ê =  1000 50000 1000 , t = 0.1, F = 10,
1000 1000 1000
3
X
1000 ≤ Ei ≤ 100000, 300 ≤ t λj ≤ 12000. (6)
j=1
Fig. 1 and Fig. 2 show the significant change in compliance through the process.
In the patch analysis the whole structure is divided into a fixed number of patches, in which the materi-
als parameters are the same for all elements belonging to this patch. This reduces the number of design
parameters to a much lower number, depending on the number of assumed patches. Thereby the size of
patches is the same for all of them and they are disposed in a structured mesh (cf. Fig. 3).
Figure 3. Subdivision of area in different number of patches, with equal material properties
Table 1. Results in compliance, number of iterations and computing time with respect to number of patches
Number of patches Compliance Number of iterations Computing time in s
1 0.9687 153 47
2 0.9653 157 60
4 0.9563 178 67
8 0.8854 217 84
32 0.8471 414 162
128 0.8392 630 7462
In Tab. 1 it is shown that for a low number of patches the number of finite elements is the same for
all models. The total number of design variables can be preserved by the number of patches and 6
design variables per patch. The number of necessary iteration and the computing time does not change
285
significantly. Also, the compliance is decreasing slowly. For a very high number of patches, the number
of iteration and most strikingly the computing times are increased significantly.
Acknowledgments
The founding of this project by the German Research Foundation (DFG) and the provision of the MMA
algorithm by Krister Svanberg [2] is kindly acknowledged.
References
[1] N. Gerzen, F. J. Barthold, S. Klinkel, W. Wagner, and D. Materna. Variational sensitivity analysis
of a nonlinear solid shell element. International Journal for Numerical Methods in Engineering 96.1
(2013), pp. 29–42. doi: 10.1002/nme.4545. url: https://doi.org/10.1002%2Fnme.4545.
[2] Krister Svanberg. The method of moving asymptotes—a new method for structural optimization.
International Journal for Numerical Methods in Engineering 24.2 (1987), pp. 359–373. doi: 10 .
1002/nme.1620240207. url: https://doi.org/10.1002%2Fnme.1620240207.
[3] Alemseged Gebrehiwot Weldeyesus and Mathias Stolpe. Free material optimization for laminated
plates and shells. Structural and Multidisciplinary Optimization 53.6 (2016), pp. 1335–1347. doi:
10.1007/s00158-016-1416-3. url: https://doi.org/10.1007%2Fs00158-016-1416-3.
286

Mesh Tying and Contact Algorithms for Nonlinear Beam-

to-Solid Interactions
1
Ivo Steinbrecher ∗
and Alexander Popp1
1 Institute for Mathematics and Computer-Based Simulation, University of the Bundeswehr Munich (UniBw M), Germany
Abstract. The interaction of Rod- or Beam-like structures with 3-dimensional continua (solids) can be found in
a variety of different physical problems, e.g. fiber-reinforced materials. In many cases a complete 3-dimensional
model of the beams is numerically expensive and leads to a considerable modeling effort. A homogenization
of the composite material on the other hand, comes with limitations when detailed information about the
interaction between beams and solid is desired. This work introduces a beam-to-solid volume coupling method,
based on classical mortar methods, which allows for the efficient coupling of 1-dimensional beams models to
3-dimensional solids, while still being able to preserve their complex effect on the overall response of the
coupled structure.
Introduction
From a mechanical point of view, rod- or beam-like structures can be represented by a 1-dimensional
continuum, described by a centerline and – depending on the kinematic assumptions of the employed
beam theory – up to three rotations of the beam cross-section. This leads to a reduced set of governing
equations compared to a full 3-dimensional continuum, thus allowing for an efficient numerical analysis
of such structures.
Going beyond purely beam-based structures, many applications also require accurate models for the fully
nonlinear interaction between beams and solid bodies. Beam-to-solid interaction can be divided into three
elementary cases: beam-to-volume coupling, beam-to-surface coupling and beam-to-surface contact. In
the presented case of beam-to-volume coupling, a beam is embedded inside a solid volume, and the beam
is tied to the solid. For the numerical simulations, we utilize the finite element method (FEM) to discretize
both the beams and solids [2, 5]. The constraints are discretized with mortar coupling methods, which
have become state-of-the-art in the coupling of non matching rids on deformable solids [3, 4].
1 Problem formulation
We consider a 3D finite deformation beam-to-solid volume coupling problem as shown in Figure 1. The
principle of virtual (PVW) work serves as basis for the employed finite element method. Contributions
to the total virtual work of the system can be split up into solid, beam and coupling terms, whereat the
beam and solid terms are independent the coupling constraint. Therefore well established formulations
for the solid as well as the beam can be used without modification. Obviously, this introduces a modeling
error, since we assume overlapping material points, but nonetheless this allows for an efficient and simple
model of the solid, while still being able to account for complex nonlinear effects due to the embedded
beams.
The beam-to-solid coupling constraints are formulated on the beam surface ΓB in the reference configu-
ration and read
uB − uS = 0 on ΓB . (1)
The mortar method is employed to discretize the coupling, introducing a Lagrange multiplier vector field
λ defined on ΓB , which can be interpreted as the negative interface tractions acting on the beam surface.
∗ Corresponding author: ivo.steinbrecher@unibw.de
287
ΩB
0
ΩS0
uB
XB ΩSt
XS
e3 uS
e2
e1
ΩB
t
Figure 1. Notation of the finite deformation beam-to-solid volume coupling problem.
Contributions to the total virtual work are the virtual work of the coupling tractions and the variational
form of the coupling constraints
Z Z

−δWcf = λ δuB − δuS dA0 , δWλf = δλ uB − uS dA0 . (2)
ΓB ΓB
These integrals are evaluated on the coupling surface ΓB which results in computationally expensive
numerical integration. If the cross-section radius of the beams is small compared to the other dimensions
of the beam-to-solid volume coupling problem, we can approximate those surface integrals as line integrals
along the beam centerline
Z Z
f B S
f
δWc ≈ λ δu − δu ds = δWc , δWλ ≈ δλ uB − uS ds = δWλ . (3)
L L
This approximation changes the coupling applied on the physical beam-to-solid volume coupling model.
The full surface integrals represent a surface-to-volume coupling, while the approximated line integrals
model a line-to-volume (1D-3D) coupling. This is a significant change in the mathematical description
of the mechanical model, which has implications on the coupling of beam cross-section rotations. The
coupling terms are now exclusively formulated along the beam centerline, which uncouples the cross-
section rotations from the solid deformations.
2 Spatial discretization and numerical integration
A finite element discretization is employed to approximate the continuous fields for position, displacement
and Lagrange multipliers, as well as their variations. The discretization introduces discrete values for the
field functions and interpolates them with corresponding shape functions,
nS
X nB
X nλ
X
uSh = Nk dS , uB
h = Hk dB , λh = Φk λ. (4)
k=1 k=1 k=1
For the interpolation of the Lagrange multipliers standard linear Lagrangian shape functions are em-
ployed. Inserting the finite element discretization into the virtual work of the coupling terms leads to the
discretized coupling interactions
T T
−δWc,h = δdB D T λ − δdS M T λ (5)
T B T S
δWλ,h = δλ Dd − δλ M d , (6)
where D and M are the so called mortar coupling matrices defined as

Z Z
D[j, k] = Φj Hk ds I, M [j, k] = Φj (Nk ◦ χh ) ds I. (7)
L L
All integrals are evaluated on the beam centerline, therefore a projection of the solid shape functions onto
the beam centerline is required, which is denoted with the nonlinear mapping operation Nk ◦χh . Accurate
288
numerical integration is fundamental to fulfill basic consistency tests for beam-to-solid volume coupling.
Figure 2 illustrates the two different types of numerical integration schemes used to evaluate the coupling
terms. Element-based integration uses a fixed number of Gauss–Legendre integration points per beam
element, only if a beam sticks out of the solid domain (strong discontinuity), the integration domain is
adapted. Segment-based integration on the other hand also splits the integration domain at points where
the beam crosses between two solid elements (weak discontinuity). A comprehensive comparison between
the different numerical integration schemes in the context of 3-dimensional mortar coupling can be found
in [1].
element-based integration segment-based integration
beam node segmentation point Gauss point
Figure 2. Illustration of element-based (left) and segment-based (right) integration.
The mortar coupling constraints are enforced via the penalty method, resulting in the following global
quasi-static equilibrium equations
S S
K SS + M T M −M T D d r
= B , (8)
−D T M K BB + D T D dB r
where is the penalty parameter.
3 Examples
Figure 3 shows the patch test setup used to verify the consistency of the beam-to-solid volume coupling
method, as well as the results obtained with different numerical integration schemes for the coupling
terms. It is clear that segment based integration gives better results than the element based integration,
i.e. a lower integration error.
Figure 3. Patch test with overlapping straight beams – problem setup (left), element-based integration (middle)
and segment-based integration (right). The Cauchy stress σzz is shown in the solid and the curvature κ
at the middle of each beam element. Displacements of beams and solid are scaled with a factor of 100.
The spatial convergence of the beam-to-solid volume coupling method is shown in Figure 4. One can see
that for a certain element size the approximation of the coupling terms as line integrals results in a non
289
converging behavior. At this point the solid element size is in the same range as the beam cross-section
dimensions and the line integral approximation does not accurately describe the coupling terms.
10−2 O(h2el )
surface integration
10−3 mortar line2
eL2
10−4
10−1 100
hel
Figure 4. Spatial convergence behavior of the displacement error kekL2 – the solid mesh with the beam cross-
section is shown for the mesh configuration where the convergence of the centerline coupling stops.
To demonstrate the applicability of the beam-to-solid volume coupling method to problems with fiber-
reinforced materials, Figure 5 shows a composite plate with an unsymmetrical layer setup, where the
fibers are modeled using beam elements and the matrix is modeled with solid-shell elements. In the right
part of Figure 5 the coupling tractions between the fibers and the matrix are shown.
Figure 5. Composite plate, fibers modeled with beam elements – deformed structure (left), coupling tractions
between the beams and the matrix (right).
References
[1] P. Farah, A. Popp, and W. A. Wall. Segment-Based vs. Element-Based Integration for Mortar Meth-
ods in Computational Contact Mechanics. Computational Mechanics 55.1 (2015), pp. 209–228.
[2] C. Meier, A. Popp, and W. A. Wall. Geometrically Exact Finite Element Formulations for Slender
Beams: Kirchhoff–Love Theory Versus Simo–Reissner Theory. Archives of Computational Methods
in Engineering 26.1 (2017), pp. 163–243.
[3] A. Popp and P. Wriggers. Contact Modeling for Solids and Particles. Springer, 2018.
[4] M. A. Puso. A 3D Mortar Method for Solid Mechanics. International Journal for Numerical Methods
in Engineering 59.3 (2004), pp. 315–336.
[5] P. Wriggers. Nonlinear Finite Element Methods. Springer-Verlag Berlin Heidelberg, 2008.
290

Analysis of support conditions applying coupling-based

deformation modes in Generalised Beam Theory (GBT)
Fabiola Tartaglione
∗ 1, 1 1
Marcelo J. Bianco , Abinet K. Habtemariam , Volkmar Zabel
1 and Carsten Könke
1
1 Institute of Structural Mechanics. Bauhaus-Universität Weimar, Germany
Abstract.
Thin-walled structures are generally subjected to mechanical eects such as cross-sectional ovalisation and
warping. Existing approaches to study their mechanical behaviour require computationally intensive shell
element models. The Generalised Beam Theory (GBT) is a numerical method that has an excellent perfor-
mance for analysing local and global stability of thin-walled structures, overcoming the computational eorts
demanded by shell element models. While conventional GBT analysis has made substantial progress, the
coupling of GBT deformation modes due to partial support conditions has not been addressed yet. This
research focuses on the implementation of multi-freedom constraints in linear GBT for modelling the existing
coupling between deformation modes due to sti or exible spring support conditions. The applicability of the
proposed method is veried through numerical simulations of an above-ground pipeline system with various
partial support conditions and compared with Kircho-Love and Reissner-Midlin shell models. The results
demonstrated accurate predictions in the displacement and stress elds using the GBT coupling analysis. It is
also concluded that increasing the number of participating deformation modes in the GBT analysis revealed
a higher correlation of the results with respect to the shell FEM solutions.
Keywords: above-ground pipeline systems, generalised beam theory, master-slave method, coupling defor-
mation modes, exact nite element formulation, partial supports.
Introduction
Above-ground pipelines are thin-walled structural systems that are generally analysed using shell ele-
ments. Nevertheless, the use of these elements would result in large computational cost due to the large
amount of degrees of freedom (DOF) needed. The detailed mechanics that must be described requires
also rigorous and precise treatment. Therefore, a remarkable alternative for modelling above-ground
pipelines is the Generalised Beam Theory (GBT).
GBT is a state-of-the-art approach which analyses local and global stability behaviour of thin-walled
members by incorporating 2D aspects of shell theories into the classical beam theory, providing the fac-
ulty to describe distortion and shear deformation [6]. Although the theory provides accurate results by
itself [2, 4, 6], further developments in terms of the coupling between deformation modes have not been
addressed yet.
Typically, boundary conditions in GBT application are assumed orthogonal for several engineering prob-
lems. However, non-standard support conditions can no longer sustain this major assumption, as the
boundary conditions act as a coupler factor among the traditional uncoupled GBT modes. Non-standard
boundary conditions of above-ground pipelines, namely partial supports over the perimeter of the cross-
section, lead to coupling of GBT deformation modes.
In this work, the implementation of multi-freedom constraints (MFC) in linear GBT is carried out to
model the existing coupling between deformation modes due to partial localised sti or exible spring
support conditions. GBT elements are implemented based on an exact nite GBT element formula-
tion [1], while MFC are based on the master-slave method, which serves as a coupling relation between
all GBT deformation modes. The applicability of the coupling method is veried through numerical
simulations of an above-ground pipeline system with various partial support conditions and compared
with Kircho-Love and Reissner-Midlin shell models.
∗ Corresponding author: fabiola.tartaglione@uni-weimar.de ORCID iD: https://orcid.org/0000-0001-6566-4682
291
1 GBT analysis
GBT is a numerical method introduced by Schardt [5] that adds to the kinematics of Vlasov the distortion
of the cross-section by means of expressing the deformation of the member as the superposition of a
series of cross-sectional modes. Hence, GBT equations are decoupled with respect to normal stresses
and transverse stresses [4]. GBT's main feature is the cross-section analysis, which is based on the
geometry and mechanical properties of the cross-section. The analysis leads to n natural orthogonal
deformation mode shapes, where lower modes (i=1,2,3) represent the rigid body modes, i.e. elongation
and bending in both transverse directions; and higher modes (i>3) describe the cross-section distortions
and ovalisations. The analysis for a circular hollow sections (CHS) is dened by orthogonal deformation
modes (see Figure 1a) based on Fourier series [5]. The displacement eld is described for the centre line
of the cross-section at each angular position θ as:
n
X n
X n
X
i
u (x, θ) = u (θ) i V,x (x) v (x, θ) = i
v (θ) i V (x) w (x, θ) = i
w (θ) i V (x) (1)
i=1 i=1 i=1
where u, v and w are the displacements in longitudinal, transverse tangential and transverse perpendicular
i
directions, respectively; while V (x) is the amplication function along the beam length for each GBT
i
deformation mode i. In the longitudinal displacement u, the rst derivative V,x (x) grants the possibility
to describe zero-shear deformations.
Figure 1. (a) Cross-section deformation mode shapes according to GBT.

(b) Partial supports types for above-ground pipeline systems.
2 Coupling-based deformation modes in GBT
One of GBT's main assumptions is the orthogonality between modal displacements and stress elds.
Thus, the nal GBT equations are decoupled with respect to normal stresses and transverse stresses
[3]. Although typical boundary conditions - Dirichlet or Neumann types - are assumed to be orthogonal
in GBT, non-standard support conditions such as the partial nodal supports in Figure 3b, suggest a
factorisation between the conventional GBT deformation modes. Hence, the nal displacement state of a
cross-section is a relation between the rigid-body modes and the ovalisation modes, which is found at each
support node. The MFC is carried out by means of the master-slave method, which consists in dening
the master and slave DOF. The slave DOF are then explicitly eliminated. For the implementation of the
method, the boundary condition can be expressed as:
n
sp
X i

δw θ̄ = w θ̄ i V (x = xsp ) (2)
i=1
sp
where w is the unknown spring displacement located at θ̄. This displacement shall be equal to the
summation of all GBT participating modes, which are aected by the amplication function at the
longitudinal position of the support xsp . From this expression a coupling vector is obtained, which is
represented by the slave mode as function of the remaining deformation modes. Therewith, the modied
displacement can be determined and transformed into the actual displacement. A scheme of this procedure
is presented in Figure 2.
Two general conditions arise in the analysis: (i) each spring support leads to one coupling equation which
require at least two GBT deformation modes to be coupled; as a consequence, (ii) GBT deformation
292
Figure 2. Scheme of multi-freedom constraints (MFC) application.
modes with and without external loads have a contribution to the total beam stiness. In order to satisfy
the conditions, for n coupling springs, the number of GBT modes to consider shall be n + 1. Additionally,
due to symmetric load conditions as illustrated in Figure 3b, the actual presence of locking eect can be
avoided by considering only half of the total of spring equations for the complete cross-section analysis.
3 Numerical example
The coupling method is veried through a symmetric 4-span above-ground pipeline system. The partial
nodal supports shown in Figure 3b have been dened -in structural terms - as elastic translational springs
with a stiness of 500N/mm. The pipeline system is subjected to self-weight and a linear projected surface
load due to gas weight. The material properties are Young's modulus E = 2.1 · 105 MPa and Poisson's
ratio ν =0.3.
Figure 3. (a) Longitudinal discretisation in GBT ; (b) Cross-section and external forces ;
(c) Internal forces orientation
The cross-section deformation and internal forces obtained from GBT and FEM shell analysis for section
A-A are presented in Figure 4 and 5. Considering 7 coupling equations, the number of GBT modes to
analyse is then n = 8. Two analyses in GBT for this case were conducted. The rst analysis involves only
conventional deformation modes, while in the second analysis 28 modes including the extension mode
(i = 1) are assessed. As far as shell analysis is regarded, linear Kircho-Love and Reissner-Midlin shell
R were considered.
elements in Ansys
Figure 4. (a) Cross-section deformation ; Longitudinal and transverse (b)membrane forces (c) plate forces
293
Figure 5. (a) Longitudinal, transverse and torsion moments; (b) Shear membrane force
Regarding the numerical performance, the required amount of DOF in the GBT analysis was around
544, compared to the nearly 1.44 Million DOF required in the shell analysis. This comparison evidences
one of the most remarkable aspects from GBT analysis. Results from GBT regarding the displacement
eld at support nodes in section A-A showed an excellent agreement with shell element models. Also,
when more than the minimum required GBT modes were taken into account, the correlation between
GBT and shells tended to increase. In the stress eld, results from both analyses showcased a disruption,
which can be associated to the implementation of non-standard support conditions. Therewith, results
of internal forces illustrated in Figure 4 correspond to a cross-section B-B, which is at 0.5m distant from
section A-A that presented consistent results for both analyses. The results obtained in GBT have also
shown a high correlation with respect to shell analysis. As seen in the displacement eld, correlation
between models increases when more GBT deformations modes are included.
4 Conclusions
In this work, the implementation of MFC in linear GBT to model the coupling between deformation
modes due to non-standard support conditions has been conducted. The Master-Slave method has been
used due to its straightforward implementation. With the application of this coupling method, main
assumptions from conventional GBT have been changed. Therewith, the new linear equation system
must be solved considering that the coupling equations generated due to boundary conditions requires
a minimum amount of deformation modes. Consequently, all deformation modes - with and without
external loads - are coupled and contribute to the internal stress energy. Results have demonstrated
that GBT presents an outstanding numerical performance with respect to shell analysis. Moreover, an
excellent correlation between shell models and GBT has been obtained for both displacement and stress
elds for several cases. A dierence between results can be diminished if more GBT deformation modes
are included in the GBT analysis for non-standard support conditions.
Acknowledgments
This work was carried out under the Horizon 2020 MSCA-RISE-2015 project No. 691213 entitled Ex-
change Risk.
References
[1] M.J. Bianco, C. Könke, A.K. Habtemariam, and V. Zabel. Exact nite element formulation in gen-
eralized beam theory. International Journal Advanced Structural Engineering (2018), 295323.
[2] D. Camotim, N. Silvestre, C. Basaglia, and R. Bebiano. GBT-based buckling analysis of thin-walled
members with non-standard support conditions. Thin-Walled Structures 46 (2008), pp. 800815.
[3] J. Jönsson and M.J. Andreassen. Distortional eigenmodes and homogeneous solutions for semi-
discretized thin-walled beams. Thin-Walled Structures 49 (2011), pp. 691707.
[4] J. Jönsson and M.J. Andreassen. Distortional solutions for loaded semi-discretized thin-walled beams.
Thin-Walled Structures 50 (2012), pp. 116127.
[5] Richard Schardt. Verllgemeinerte Technische Biegetheorie. Springer-Vertrag, 1989.
[6] N. Silvestre and D. Camotim. Inuence of shear deformation on the local and global buckling be-
haviour of composite thin-walled members. In: Proceedings of the 4th International Conference on
Thin-Walled Structures. Ed. by J. Loughlan. IOP Publishing Ltd, 2003, pp. 659668.
294
MS 16: Recent Advances on Time Integration with Focus on CFD Applications

Multirate GARK schemes for the simulation of fluid-

structure interaction: Linear stability and the geometric
conservation law
Sascha Bremicker-Trübelhorn∗1 and Sigrun Ortleb1
1 University of Kassel, Department of Mathematics, Heinrich-Plett-Str. 40, 34132 Kassel, Germany
Abstract. A novel loosely coupled partitioned approach to the simulation of mechanical fluid-structure in-
teraction (FSI) problems based on step size-adaptive multirate GARK schemes (MGARK) and discontinuous
Galerkin schemes (DG) is presented. We use mixed implicit-explicit (IMEX) GARK schemes of orders up to
three for the temporal integration of the DG-discretized partitioned fluid-structure system and apply these
GARK schemes in a multirate fashion to respect the different time scales of the involved subsystems. An
efficient coupling between the subsystems is realized by using a compound step-based coupling strategy that
preserves the order of the base RK schemes. The linear stability properties of the overall temporal integra-
tion scheme are investigated and it is shown how the geometric conservation law (GCL) can be respected by
calculating the velocities of the moving mesh points in an appropriate way. Additionally, we consider the rela-
tion between geometrical conservation and the ability of the scheme to preserve its order obtained on a fixed
mesh also in the presence of arbitrary mesh motion. The feasibility and efficiency of our approach is investi-
gated by applying it to a two-dimensional test case of inviscid compressible flow around an oscillating airfoil.
Introduction
Coupled systems as in the context of fluid-structure interaction (FSI) often consist of subsystems with
significantly different time scales. These subsystems may also deviate considerably in terms of their stiff-
ness. It is reasonable to allow each subsystem of the coupled problem to advance with its preassigned
time integration scheme which is adapted to its stiffness and time scales. In addition, it is sensible to
evaluate slow components less often than fast ones. In [1, 2] the authors have shown how the class of
multirate generalized-structure additively partitioned Runge-Kutta schemes [3]—the multirate general-
ization of the class of GARK schemes [6]—can be used to do so, and successfully applied it in a novel
step size-adaptive fashion to the 1D piston problem, a classical test problem of mechanical FSI.
This paper shows the feasibility of this approach for the application to a more complex 2D problem of
mechanical FSI and is structured as follows: First, in Sec. 1 we briefly describe the considered test case
and its spatial discretization. In Sec. 2 we sketch our multirate time integration approach and investigate
its linear stability properties. Sec. 3 gives insight on how to respect the geometric conservation law (GCL)
within our 2D-DG-MGARK approach and on the relation of GCL-compliance and the preservation of our
scheme’s temporal order of convergence in the presence of mesh motion. The numerical results obtained
from applying our scheme to the described test case of FSI are shown in Sec. 4. The paper finishes with
a summary of our conclusions in Sec. 5.
1 Test case of fluid-structure interaction and its spatial discretization
We consider an oscillating NACA0012 airfoil with two degrees of freedom—which are rotation and vertical
displacement—that is exposed to transonic compressible fluid flow with a Mach number of 0.8. This
problem has already been used in the past to study the performance of multirate IMEX procedures [5].
The fluid flow is considered to be inviscid and is therefore modeled with the compressible Euler equations.
The airfoil is modeled as an harmonic oscillator driven by the external forces lift and torque, that are
exerted by the fluid surrounding it.
∗ Corresponding author: struebel@mathematik.uni-kassel.de
295
The 2D fluid equations are spatially discretized within the arbitrary Lagrangian-Eulerian (ALE) frame-
work using a discontinuous Galerkin (DG) approach on a triangle mesh comprising 9050 triangular cells.
In order to take into account the mesh motion as a result of the deformation of the spatial fluid domain
due to the structure’s movement, radial basis function interpolation is used to determine the inner mesh
nodes with respect to the current state of the structure system.
2 Multirate generalized-structure additively partitioned Runge-Kutta schemes
We regard the problem described in Sec. 1 as a partitioned system of first-order ordinary differential
equations, which comprises two components: the DG-discretized fluid component and the structure
component, which corresponds to the oscillating airfoil. Due to its generally shorter time scales, the fluid
subsystem is considered as the fast component and is advanced with several small and generally adaptive
so-called micro steps hλ , while at the same time the slow component (i. e. the structure) is advanced
with only one large so-called macro step H. Let N denote the ratio of micro steps per macro step, which
is allowed to change from one macro step to the next due to the adaptivity of the step sizes. With
mλ = H/hλ (λ = 1, . . . , N ) the generalized Butcher table of our scheme reads
m1 A{f,f} A{f,s,1}
m1 1b{f}T m2 A{f,f} A{f,s,2}
{1,1} {1,2}
A A .. .. .. ..
{2,1} {2,2}
. . . .
A A =
m1 1b{f}T m2 1b{f}T ··· mN A{f,f} A{f,s,N }
b{1}T b{2}T
m1 A{s,f,1} 0 ··· 0 A{s,s}
.
m1 b{f}T m2 b{f}T ··· mN b{f}T b{s}T
Since only the very first micro step is coupled to the macro step, all further coupling coefficients of the
fluid-to-structure-coupling part of the Butcher table are set to zero. We chose explicit RK base schemes
(A{f,f} , b{f} ) for the integration of the fluid subsystem and implicit RK base schemes (A{s,s} , b{s} ) for
the integration of the structure subsystem. Further details on our order-preserving fully explicit coupling
strategies and on the applied IMEX base schemes can be found in [2].
2.1 Linear stability
We studied the linear stability properties of our schemes using the two-times component partitioned
test equation u0 (t) = Eu(t), with u : R≥0 → R2 , E ∈ R2×2 . Let H denote the macro step size and
Z = HE. We computed the amplification matrix resulting from applying our micro step size-adaptive
MGARK scheme to the test equation and showed that its eigenvalues can be expressed in terms of the
three quantities
z1,1 z1,2 z2,1 z2,2
ξ= , η= and κ= ,
1 − z1,1 2z1,1 z2,2 − z1,2 z2,1 z1,1
which correspond to the admissibly macro step size, strength and orientation of the coupling, and the
stiffness ratio of the subsystems, respectively. We showed that in the case of the invertibility of the
matrices I − z1,1 A{1,1} and I − z2,2 A{2,2} − z1,2 z2,1 A{2,1} (I − z1,1 A{1,1} )−1 A{1,2} the subset
z z2,1
Z ∈ R2×2 z1,1 , z2,2 < 0 ∧ z1,2
1,1 z2,2
6= 2 ∧ ∀λ ∈ σ(Z) : Re(λ) < 0 ∩ S
of the whole stability region S can be expressed in terms of (ξ, η, κ) ∈ (−1, 0) × (−1, 1) × (0, ∞), which
on the one hand provides a more compact representation of that region and on the other hand reduces
its dependence from four variables to three. By numerically determining this stability region’s subset in
terms of ξ, η and κ for the developed MGARK schemes of orders up to three and a varying ratio of micro
steps per macro step, we found that our first order scheme has a large stability region even in the case
of high step size rates N and turned out to be well-suited for the application to stiff problems. However,
in the case of high stiffness the third order method suffers from a severe loss of stability under increasing
N , which makes it less suited for the efficient application within our multirate framework.
296
3 The geometric conservation law
For numerical simulations involving the movement of the spatial mesh, special attention on geometrical
conservation has to be paid. The usage of the ALE-formulation for the spatial discretization of the fluid
equations led to a modification of the convective flux tensor depending on the velocities of the moving
mesh nodes. Our schemes satisfy the geometric conservation law (GCL)—i. e. they exactly preserve
any initially uniform flow—by the careful construction of these mesh velocities within every RK-stage of
every micro step of the fast fluid RK-integration in a suitable way. To do so, we extended an approach
derived in [4] to the application to our 2D-DG-MGARK schemes. The approach is based on a space-time
DG-discretization of the fluid equations, which is known to automatically satisfy the GCL, and a linear
and continuous finite element ansatz for the mesh velocity. The equations for the mesh velocity obtained
within this space-time DG-formulation are then translated into the DG-MGARK context.
We verified the GCL-compliance of our schemes of orders up to three both analytically and numerically,
using an artificial mesh motion on a 2D test grid. It was observed that a naive choice of the mesh velocity
by using its analytical values obtained by differentiating the artificial mesh motion equations led to a
violation of the GCL. Nevertheless, the GCL-compliant choice of the mesh velocity led to a loss of the
temporal order of convergence for the third order MGARK scheme in the case of simulations with a
non-uniform initial condition. Since this loss of order was not observed with the naive and non-GCL-
compliant choice of the mesh velocity for this kind of simulations, we conclude that GCL-compliance is
neither a sufficient nor necessary condition for a numerical scheme to preserve its order of convergence
obtained on a fixed mesh also in the presence of arbitrary mesh motion.
4 Numerical results for the considered test case
The results shown in Fig. 1 were obtained by applying the GCL-compliant DG-MGARK schemes of
orders one and two to the test case described in Sec. 1, using constant step sizes and various ratios N
of micro steps per macro step. In each case the spatial order of the DG-approximation is chosen to
match the temporal order of the respective MGARK scheme. The reference solution was computed with
a monolithic all-explicit RK procedure of the respective order of convergence for the integration of the
overall FSI system using a suitable small uniform time step size.
Reference N =1 N =5 N = 10 N = 20 N = 40 Reference N =1 N =5 N = 10 N = 20 N = 40
Time t̂ [ s ] Time t̂ [ s ]
Angle of rotation θ̂ [ ◦ ]
Angle fo rotation θ̂ [ ◦ ]
0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20
0.5 0.5
0.0 0.0
−0.5 −0.5
p(x) p(x) p(x) p(x) p(x) p(x)

Displacement ŷ [ cm ]
Displacement ŷ [ cm ]
5.0 5.0
0.0 0.0
−5.0 −5.0
0.00 0.05 0.10 0.15 0.00 0.05 0.10 0.15

Time t̂ [ s ] Time t̂ [ s ]
Figure 1. Evolution of the angle of rotation θ̂ and of the vertical displacement ŷ of the NACA0012 airfoil for
different step size rates N . The fluid pressure p in the environment of the airfoil is shown at three
selected points in time. Left: First-order simulation. Right: Second-order simulation.
For step size ratios up to N = 20, the results obtained with the first-order scheme are close to or very
close to the reference solution. Only the case N = 40 shows clearly recognizable deviations from the
reference in terms of the amplitude and frequency of the observed oscillation. The second-order scheme
shows very similar behavior with regard to the quality of the results obtained for up to and including
297
N = 20. The results for the case N = 40, however, show a superposition of the actual oscillation with
a higher frequency and unphysical disturbance. This loss of stability observed for higher ratios of step
sizes becomes even more severe for the third-order scheme (not shown in the figure), in case of which
the simulations with N ≥ 20 diverged after a short time and N = 5 resp. N = 10 produced slight resp.
strong deviations from the reference oscillation.
Therefore, the first- and second-order scheme proved to be well suited for the application within the
described multirate context in case of not too large step size ratios, whereas the third-order scheme
turned out to be less suited for this kind of application.
5 Conclusion
This paper focused on the linear stability properties and the compliance with the geometric conservation
law of our loosely coupled step size-adaptive multirate partitioned mixed implicit-explicit approach [1, 2]
to the simulation of mechanical fluid-structure interaction. We showed the feasibility of combining DG-
based high-order spatial discretization with high-order MGARK-based temporal discretization schemes
and successfully applied our scheme to the simulation of 2D transonic compressible flow around an
oscillating airfoil. The GCL-compliance of the scheme is assured by the careful construction of suitable
mesh velocities within every RK-stage of every micro step of the fluid RK-integration. Both the linear
stability analysis and the numerical results showed that the first- and second order schemes are well suited
for the application within the considered multirate context, whereas the third-order scheme suffers from
severe stability issues. Future research might therefore focus on the construction of a more robust and
stable third-order DG-MGARK scheme.
References
[1] S. Bremicker-Trübelhorn and S. Ortleb. Time-adaptive multirate generalized-structure additively
partitioned Runge-Kutta schemes for mechanical fluid-structure interaction. PAMM 16.1 (2016),
pp. 869–870. doi: 10.1002/pamm.201610423.
[2] S. Bremicker-Trübelhorn and S. Ortleb. On Multirate GARK Schemes with Adaptive Micro Step
Sizes for Fluid–Structure Interaction: Order Conditions and Preservation of the Geometric Conser-
vation Law. Aerospace 4.1 (2017). doi: 10.3390/aerospace4010008.
[3] M. Günther and A. Sandu. Multirate generalized additive Runge Kutta methods. Technical Report
CSL-TR-6/2013. Computational Science Laboratory, 2013. arXiv: 1310.6055.
[4] D. J. Mavriplis and C. R. Nastase. On the geometric conservation law for high-order discontinuous
Galerkin discretizations on dynamically deforming meshes. J. Comp. Phys. 230.11 (2011), pp. 4285–
4300. doi: 10.1016/j.jcp.2011.01.022.
[5] S. Piperno. Explicit/implicit fluid/structure staggered procedures with a structural predictor and
fluid subcycling for 2D inviscid aeroelastic simulations. International Journal for Numerical Methods
in Fluids 25 (1997), pp. 1207–1226.
[6] A. Sandu and M. Günther. A Generalized-Structure Approach to Additive Runge-Kutta Methods.
SIAM J. Numer. Anal. 53.1 (2015), pp. 17–42.
298

Stability Analysis of IMEX-DG-Schemes for Advection-

Diffusion with Focus on the Diffusion Treatment
Sigrun Ortleb
FB Mathematik und Naturwissenschaften, Universität Kassel; ortleb@mathematik.uni-kassel.de
Abstract. This talk describes a fully discrete L2 -stability analysis for linear 1D advection-diffusion problems,
discretized in space by the DG scheme based on the (σ, µ)-family of diffusion discretizations and implicit-
explicit Runge-Kutta schemes in time. Advection terms are discretized explicitly in time while diffusion terms
are solved implicitly. Conditions on the parameters σ and µ are derived which guarantee L2 -stability for
time steps ∆t = O(d/a2 ), where a and d denote the advection and diffusion coefficient, respectively, i.e. the
allowable time step size does not decrease under
grid refinement. These conditions are fulfilled in particular by
the BR2 scheme as well as the recent 41 , 49 -recovery scheme. However, the BR1 scheme and the Baumann-
Oden method do not possess a similar stablility property which is also confirmed by numerical experiments.
Introduction
In computational fluid dynamics, advection-diffusion problems usually arise when the time evolution
of chemical substances or biological species in a flowing medium is described. Furthermore, the flow of
viscous fluids itself, modelled by the compressible or incompressible Navier-Stokes equations contains both
convective and viscous terms which have different characteristics when discretized in space. In particular,
the resulting systems of ordinary differential equations become increasingly stiff under grid refinement due
to the presence of the diffusion terms. Explicit time integration then leads to severe time step restrictions
which reduce the time step with the square of the grid size. Purely implicit time discretization requires
the solution of large non-linear systems of equations. This has lead to the development of implicit-explicit
(IMEX) methods to be applied to advection-diffusion type problems. Hereby, a common approach is to
discretize advection terms explicitly while the diffusion terms are treated implicitly. While this approach
alleviates the severe time step restriction caused by the diffusion terms, a CFL-type time step restriction
may still have to be fulfilled due to the explicit treatment of the advection terms, scaling the time step
with the grid size. However, for some combinations of space and time discretization schemes, the time
step does not have to be reduced under grid refinement. This favourable stability property has already
been shown in the literature for specifically designed IMEX methods and Fourier spatial discretization,
where the time step solely depends on the advection and diffusion coefficients. A similar unconditional
stability result has recently been proven by Wang et al. [2] in the context of spatial discretization by the
local discontinuous Galerkin (LDG) method. Since Wang et al. consider a specific diffusion discretization,
the question arises if this favourable property is also inherent to more general DG methods. In this talk,
conditions on the parameters of a larger family of DG diffusion schemes are derived which allow for this
type of unconditional stability. It turns out that these parameter restrictions are neither fulfilled by
the simple first approach of Bassi and Rebay nor by the classical Baumann-Oden scheme, hence these
schemes do not allow for a grid-independent time step choice. However, both the second approach of
Bassi and Rebay with an adjusted penalty constant and the more recent recovery scheme by van Leer
and Nomura fulfill the presented conditions.
1 The IMEX-DG scheme for 1D linear advection-diffusion
We consider the linear advection-diffusion equation
Ut + aUx = dUxx , (x, t) ∈ Q = Ω × (0, T ), Ω = (xa , xb ) (1)
299
with diffusion coefficient d > 0 and advective velocity a > 0, supplemented by the periodic initial
condition U (x, 0) = U0 (x) in L2 (Ω) and periodic boundary conditions. The domain Ω is partitioned into
cells Ij = (xj− 12 , xj+ 12 ), j = 1, . . . , K with x 12 = xa , xK+ 21 = xb . Here, we consider uniform grids with
cell length xj+ 12 − xj− 12 = ∆x. The basis functions and test functions used to define the DG scheme are
taken from the finite element space
Vh = {v ∈ L2 (Ω) | v|Ij ∈ PN (Ij ) ∀j = 1, . . . , K}, (2)
where PN (Ij ) denotes the space of polynomial functions on Ij of degree at most N . As usual in DG
schemes, the functions in Vh may be discontinuous across element boundaries. At each element boundary,
the left-hand side and right-hand side values of a piece-wise continuous function v are denoted by v − and
v + , respectively. Periodicity is realized by setting v −
1 = v
−
K+ 1
+
and vK+ +
1 = v 1 . The corresponding jump
2 2 2 2
at element interfaces is denoted by [v] = v + − v − and the arithmetic mean is given by {v} = 21 (v + + v − ).
The semi-discrete DG scheme for (1) is defined as the solution u(t) ∈ Vh of the variational formulation
(ut , v)j = aHj (u, v) + dLj (u, v) , ∀t > 0, v ∈ Vh , (3)
where (·, ·)j denotes the usual inner product in L2 (Ij ) and the advective and diffusive fluxes are discretized
by the upwind numerical flux for advection resulting in
Hj (u, v) = (u, vx )j − u− v − + u−
j+ 1 j+ 1
v+ ,
j− 1 j− 1
(4)
2 2 2 2
and the (σ, µ)-family for diffusion also considered in [1], i.e.
− +
Lj (u, v) = −(ux , vx )j + {ux }j+ 12 vj+ 1 − {ux }j− 1 v
2 j− 2
1
2

σ µ (5)
+ vx− [u] 1 + vx+ [u] 1 + [u]v − 1 − [u]v + 1 .
2 j+ 2 j− 2 ∆x j+ 2 j− 2
The DG diffusion discretization is hence specified by the two parameters σ, µ ∈ R.

Now, implicit time discretization for the diffusion term and explicit time discretization for advection
yields fully discrete IMEX-DG schemes. For time stepping, two time integration methods are considered.
First order IMEX time integration yields the fully discrete scheme
(un+1 , v)j = (un , v)j + ∆t aHj (un , v) + ∆t dLj (un+1 , v), ∀v ∈ Vh , (6)
where u , u
n n+1
denote the approximations at times t and t n
= t + ∆t, while by second order IMEX
n+1 n
time integration using an intermediate stage u(n,1) we obtain

(u(n,1) , v)j = (un , v)j + ∆t γaHj (un , v) + γdLj (u(n,1) , v) ,

(un+1 , v)j = (un , v)j + ∆t δaHj (un , v) + (1 − δ)aHj u(n,1) , v (7)

+ ∆t (1 − γ)dLj (u(n,1) , v) + γdLj (un+1 , v) , ∀v ∈ Vh ,
√
where γ, δ are given by γ = 1 − 2
2
and δ = 1 − 2γ .
1
2 Theoretical stability results for IMEX-DG schemes with (σ, µ)-diffusion discretization
For the above schemes, we may show that the admissible time step guaranteeing L2 -stability is indepen-
dent of the mesh width. More precisely, the following results can be proven.
Theorem 2.1 Let N ∈ N0 denote the polynomial degree of the DG approximation space Vh defined in
(2). Let σ, µ ∈ R fulfill the condition
(1 − σ)2
µ̃ := < µ,
4ω0
where ω0 denotes the first weight of the Gauss-Lobatto quadrature with N + 1 nodes. Then, the IMEX-DG
scheme (6) with the (σ, µ)-diffusion operator Lj defined in (5) fulfills kun+1 k ≤ kun k if the time step is
bounded by
d
∆t ≤ 2 ,
a C
ν+µ
where the constant C is set to C = max{ µν , 1}, if σ = 1, and to C = µ−µ̃ , if σ 6= 1.
300
Theorem 2.2 Let N ∈ N0 denote the polynomial degree of the DG approximation space Vh defined in
(2). Let σ, µ ∈ R fulfill the condition
2
|1 − σ| + |1 + σ| √ββ2β
1 3
µ̃ := < µ,
4ω0
where ω0 denotes the first weight of the Gauss-Lobatto quadrature with N + 1 nodes and β1 , β2 , β3 are
defined by β1 = θ(2γ−1)+1
6 , β2 = 1+θ−γ(3+θ)
4 , β3 = γ(1+θ)
6 , for θ ∈ [0, 0.7] suitably chosen. Then, the
IMEX-DG scheme (7) with the (σ, µ)-diffusion operator Lj defined in (5) fulfills kun+1 k ≤ kun k if the
time step is bounded by the condition
( −1 )
a2 1 α12 2α22 θβ3 ν+µ
∆t ≤ min + , , C= ,
d C β1 β3 2α32 µ − µ̃
1−θ+(1+θ)γ
with α1 , α2 , α3 given by α1 = 2 , α2 = θ+1
2 , α3 = θ+1
2 (1 − γ) .
3 Numerical results
In this section, we compare the stability of IMEX-DG schemes with respect to the discretization of the
diffusion terms. We consider the exact solution U (x, t) = e−dt sin(x − at) to equation (1) in the interval
(xa , xb ) = (−π, π). The advection-diffusion problem is discretized in space by the second order nodal
DG scheme on Gauss-Lobatto nodes, hence the polynomial degree is N = 1. As in the theoretical
investigation, advection terms are discretized by upwind fluxes and diffusion terms are discretized either
by the LDG scheme or by various members of the (σ, µ)-family. First and second order IMEX schemes
are considered for time integration.
From the theoretical analysis, we expect the schemes to be stable for time steps ∆t ≤ τad2 , where τ is some
constant independent of grid refinement. Tables 1 and 2 show the analysis of the maximum stable time
step for different advection and diffusion parameters a and d, where we vary the number of cells K and
compute the numerical solution until the final time T = 1000. The maximum stable time step ∆tmax is
determined as the maximum time step for which the L2 -norm of the numerical solution is non-increasing
2
and the corresponding values of τ = ad ∆tmax are listed. As predicted by the theoretical analysis, the
BR1 and Baumann-Oden(BO) scheme do not admit grid-independent time step sizes. Here, we clearly
observe the behavior τ = O(∆x) for the admissible time step, analogous to the time step restrictions for
explicitly discretized advection equations. For second order IMEX time integration, further variants of
σ and µ have been tested. Here, except for the BR1 and BO scheme, the values of τ nearly coincide for
moderate sizes of the diffusion coefficient (Table 1). In addition, first order IMEX time integration yields
no restriction on the time step at all for this particular test case. For small sizes of the diffusion coefficient
(Table 2), the first order IMEX scheme yields nearly constant maximum time steps for LDG, BR2 and
the recovery scheme, whereas for second order IMEX time integration, on coarse grids, the allowable time
step size scales with the grid size. In this respect, those members of the (σ, µ)-family with a large value
of µ admit much larger time steps. On fine grids, the allowable time step sizes almost coincide again,
except for the BR1 and BO discretizations.
References
[1] B. van Leer, M. Lo, R. Gitik, and S. Nomura. A venerable family of discontinuous Galerkin schemes
for diffusion revisited. In: Adaptive High-Order Methods in Computational Fluid Dynamics. Ed. by
Z. J. Wang. World Scientific, 2011, pp. 185–201.
[2] H. Wang, C.-W. Shu, and Q. Zhang. Stability and Error Estimates of Local Discontinuous Galerkin
Methods with Implicit-Explicit Time-Marching for Advection-Diffusion Problems. SIAM J. Numer.
Anal. 53.1 (2015), pp. 206–227.
301
First order IMEX time integration

K LDG Recovery BR1 BR2 BO
20 + + 3.36e-01 + 3.85e-01
40 + + 1.62e-01 + 1.74e-01
80 + + 7.94e-02 + 8.25e-02
160 + + 3.92e-02 + 3.98e-02
320 + + 1.96e-02 + 1.96e-02
640 + + 9.26e-03 + 9.87e-03
Second order IMEX time integration
20 2.42 2.42 3.23e-01 2.45 4.50e-01
40 2.41 2.41 1.57e-01 2.42 2.35e-01
80 2.41 2.41 7.91e-02 2.41 8.89e-02
160 2.41 2.41 3.03e-02 2.41 4.01e-02
320 2.41 2.41 1.08e-02 2.41 2.05e-02
640 2.41 2.41 9.26e-03 2.41 9.87e-03
σ = −1 σ = −1 σ = 0.25 σ = 0.25 σ = 0.25
K
µ = 10 µ = 1.5 µ = 0.5 µ=1 µ = 10
20 2.45 7.52e-01 2.34 2.39 2.45
40 2.42 2.42 2.39 2.40 2.42
80 2.41 2.41 2.40 2.41 2.41
160 2.41 2.41 2.41 2.41 2.41
320 2.41 2.41 2.41 2.41 2.41
640 2.41 2.41 2.41 2.41 2.41
2
Table 1. Values of τ = ad ∆tmax , where ∆tmax is the maximum time step to ensure a non-increasing L2 -norm
for d = 0.1, a = 0.1. An entry of “+” means that for this test case, the scheme is unconditionally stable
independent of the time step.
First order IMEX time integration

20 2.15 2.09 2.20 2.15 2.05
40 2.03 2.02 2.04 2.03 2.00
80 2.01 2.01 1.57 2.01 1.66
160 2.01 2.01 7.86e-01 2.01 1.09
320 2.01 2.01 3.93e-01 2.01 4.73e-01
640 2.01 2.01 1.96e-01 2.01 2.16e-01
Second order IMEX time integration
20 6.63 6.82 6.30 6.87 6.42
40 3.45 3.78 3.15 3.80 3.21
80 1.96 2.53 1.57 2.73 1.66
160 1.40 1.77 7.82e-01 2.17 9.03e-01
320 1.50 1.48 3.92e-01 1.70 5.01e-01
640 1.42 1.42 1.96e-01 1.50 2.84e-01
σ = −1 σ = −1 σ = 0.25 σ = 0.25 σ = 0.25
K
µ = 10 µ = 1.5 µ = 0.5 µ=1 µ = 10
20 1.05e+01 6.46 6.45 6.53 1.01e+01
40 6.22 3.30 3.24 3.35 5.95
80 3.80 1.75 1.69 1.88 3.62
160 2.39 1.01 1.01 1.32 2.28
320 1.75 6.93e-01 9.38e-01 1.24 1.69
640 1.50 1.49 1.25 1.34 1.48
2
Table 2. Values of τ = ad ∆tmax , where ∆tmax is the maximum time step to ensure a non-increasing L2 -norm for
d = 0.01, a = 0.2.
302

Comparison of exponential-explicit (EXPEX) schemes

in the domain-based implicit-explicit (IMEX) setting
with IMEX-RK schemes for CFD applications
1
Veronika Straub ∗
and Sigrun Ortleb1
1 Department of Mathematics, University of Kassel, Heinrich-Plett-Str. 40, 34132 Kassel, Germany
Abstract. Computational Fluid Dynamics (CFD) problems typically require meshes, which have a low per-
centage of very small cells. For sufficiently high grid-induced stiffness, meaning having a high ratio between
the largest and the smallest cell sizes, purely implicit or purely explicit schemes may become inefficient due to
the large nonlinear and linear systems needed for implicit solvers, respectively the small timestep sizes dictated
by the smallest cell size needed for stable explicit time integration. Domain-based implicit-explicit (IMEX)
schemes turned out to outperform conventional schemes in such cases, as shown by Kanevsky, Carpenter et.
al. for IMEX-Runge-Kutta (IMEX-RK) schemes. We utilize the class of exponential integrators called EPIRK
in that domain-based IMEX-setting, resulting in schemes which we denoted as IMEX-sEPIRK. Due to the for-
mally explicit nature of the applied exponential integrators, we rename those schemes to exponential-explicit
(EXPEX) schemes. Here, we compare a first and two third order versions of it to the domain-based IMEX-
Runge-Kutta (IMEX-RK) schemes as well as to purely eplicit, implicit and exponential methods on a 2D
laminar flow in a nozzle. The EXPEX schemes prove to be competetive and even superior to other schemes.
Introduction
In CFD the inviscid flow can be modeled by the Euler-equations, which are a nonlinear system of hyper-
bolic PDE’s. The test case, we focus at is the flow in a twodimensional nozzle. The spatial domain Ω
.K
is subdivided into elements τ ∈ Th = {τi }i with Ω = ∪i τi , on which the PDE-system is discretized in
space by a Disconinuous Galerkin scheme resulting in the autonomuous, semidiscrete System of ODE’s
with given initial and boundary data:
d
U(t) = F(U(t)), U : R → RN , F : RN → RN , t ∈ [t0 , T ]. (1)
dt
The idea
.
of domain-based IMEX schemes is to split the computational domain Ω into two regions Ω =
Ωim ∪ Ωex such that the implicit region Ωim contains the smallest elements and the explicit region the
remaining. The right-hand side (abbreviated by rhs in the following) F is also split into two parts
F = FE + FI by assigning all terms linked to the explicit region to FE and the others to FI . Now, explicit
time stepping is applied to FE and an implicit (or just A- and unconditionally CFL-stable) scheme to FI
with the purpose to overcome the strict CFL-condition for explicit schemes, which dictates the maximal
stable timestep size dependent on the smallest cell size in the computational domain. Since Ωex does not
include the smallest elements of Ω any more, larger global timesteps can be chosen for both schemes,
since the implicit scheme does not underly a CFL-condition.
Our aim is to apply the exponential schemes of EPIRK type [7] in the implicit domain, since they turned
out to be more efficient than typical implicit schemes for large stiff problems [4]. We showed that they
are also unconditionally CFL-stable though their explicit nature and analyzed the linear stability and
conservation properties in [6]. In [5] we have shown that the intuitive coupling of EPIRK and ERK
schemes is limited to first order of convergence and presented the IMEX-sEPIRK schemes as a remedy
to achieve arbitrary high order schemes. Since the EPIRK schemes are explicit by definition, but are
used for the implicit part of the domain-based IMEX splitting, we rename the IMEX-sEPIRK schemes
∗ Corresponding author: vstraub@mathematik.uni-kassel.de ORCID iD: ORCID (optional)
303
to EXPEX (exponential-explicit) schemes. Nevertheless, we proceed to catogorize them in the class of

IMEX schemes, as they are applied in the domain-based IMEX-setting.
While our previous proceedings contributions concentrated on the theoretical analysis of the EXPEX
schemes, here we focus on numerical investigations and oppose them to four different types of schemes:
IMEX-RK, purely implicit, purely explicit and exponential schemes. Before that, we explain how we
devided Ω in Ωex and Ωim and demonstrate the relevance of a clever IMEX decomposition.
Numerical investigations
To examine the efficiency of the EXPEX schemes, a test case with high grid-induced stiffness in two
dimensions is taken as in [2]: The two-dimensional flow in a nozzle. This is a classical steady-state
problem for inviscid flow, where the analytical solution at the centerline y = 0 is known. Due to symmetry
of the geometry, the computational grid contains only the upper half, having the centerline of the nozzle
at the bottom of the grid (y = 0) with a symmetry condition as boundary condition. We choose inlet
and outlet data as in chapter 4.3.1 in [2], for which the flow results in a stationary normal shock located
at x ≈ 7.5623 after t > 20. To obtain a precise solution at the centerline of the shock position, the grid
is locally refined at that point Pshock = (0, 7.5623). We designed a grid with an in-house tool for grid
generation (see Fig. 2a).
Our strategy to split the domain into an implicit and explicit region by utilizing a scale parameter a
is given in Algorithm 1. The large task lies in determining a reasonable percentage of implicit cells,
i.e. finding a suitable parameter a. This effects the efficiency of the IMEX scheme to a high degree,
since too many implicit cells, on the one hand, result in larger nonlinear systems and too many explicit
cells, on the other hand, shrink the maximal stable timestep size. An intelligent IMEX decomposition
strategy is not published so far and our approach to find such an algorithm is currently in process.
Therefore, we temporary use the following experimental approach for our test case: We first analyze the
behaviour of the IMEX schemes for diverse percentages of implicit cells, starting from 0% up to ≈ 99%
with CFL adaptivity and a fixed CFL number of 0.1 until the end time t = 1. To this end, we plot
work-precision diagrams, in which we connect the points for all the different percentages of implicit cells
in ascending order for each of the IMEX schemes. The work is measured in terms of scaled right-hand
side evaluations: The number of FE -evaluations multiplied by the percentage of explicit cells plus the
number of FI -evaluations multiplied by the percentage of implicit cells, since the work for the evaluation
of the entries of FE resp. FI scales with those percentages. The precision is determined by a reference
solution for a very small timestep size. The results are shown in Fig. 1.
Determine τmax = maxτ ∈Ω |τ |;

Determine τmin = minτ ∈Ω |τ |;
τmid := τmin +τ
2
max
;
for τ ∈ Ω do
if |τ | ≤ a ∗ τmid then
τ ∈ Ωim ;
else
τ ∈ Ωex ;
end
end
Algorithm 1: IMEX decomposi-
tion strategy
Figure 1. Results for different IMEX decompositions: 0% until ≈
99% of impl. cells are connected in ascending order by
a line for each scheme. The first point of the curve (0%
of impl. cells) is that with the highest precision of the
respective scheme.
304
We see that fully explicit or mostly implicit schemes are expensive: The first and the last point of the
curve tend to the maximal computational cost. The minimum of the computational work is achieved at a
still appropriate precision. Since that minimum varies for each of the IMEX scheme, we choose the value
4.5e-02 for a resulting in 6.5% of implicit cells, for which the distance to the minimal point is relatively
close for all schemes, but gives a higher precision at a quite similar cost. We highlighted our choice of a
with a filled yellow circle in Fig. 1.
min |τ |
We define the grid-induced stiffness similar to [2] by S = minττ∈Ω
∈Ωex
|τ | with |τ | being the area of the element
im
τ . The grid-induced stiffness S for the parameter a = 4.5e-02 results in 591.55. The visualization of the
IMEX regions can be found in Fig. 2.
b) Zoom of the
a) Decomposition into an explicit region (green) and an im- implicit re-
plicit region (blue). gion.
Figure 2. IMEX region for a = 4.5e-01 resulting in 6.5% of implicit cells.
We perform computations until t = 5 with CFL-adaptive timestep sizes for CFL-numbers in [4.88e-05,
6.55e+03] starting from the largest one halfening it until the smallest one. It is known that (semi-
)explicit schemes become stable for CFL-numbers around 1 and implicit schemes are theoretically stable
for arbitrary large CFL-numbers, but the internal solvers for the nonlinear and linear systems or the
matrix exponential approximations begin to be unstable for too large timestep sizes. In the following
Table 1 we summarize the time integration schemes that are compared and their experimentally maximal
stable CFL-number.
Name Description of the scheme CFLstable

IMEX-RK 1 1st order IMEX-RK: Coupling of the implicit and explicit Euler methods 4.00e-01
EXPEX 1 1st order EXPEX method, denoted as IMEX-EPIRK 1 in [5] 8.00e-01
IMEX-RK 3 3rd order IMEX-RK method, denoted as ARK3(2)4L[2]SA in [3] 8.00e-01
EXPEX 3 3rd order EXPEX method, denoted as IMEX-sEPIRK3 in [5] 8.00e-01
EXPEX 3b 3rd order EXPEX method, denoted as IMEX-sEPIRK3b in [5] 8.00e-01
expl. Eul. the explicit Euler method 1.60e+00
expl. 2nd a 2nd order explicit RK method 1.60e+00
expl. 3rd a 3rd order explicit RK method 1.60e+00
expon. Eul. a 2nd order exponential integrator, called Epi2 in [4] 1.64e+03
EPIRK4SC a 4th order, stiffly accurate exponential integrator [4] 8.19e+02
EPIRK5C a 5th order exponential integrator [4] 1.64e+03
EPIRK5SC a 5th order, stiffly accurate exponential integrator [4] 8.19e+02
impl. Eul. the implicit Euler method >6.55e+03
ESDIRK3 a 3rd order explicit step singly diagonal RK scheme, see i.e. [1] 3.28e+03
Table 1. Considered time stepping schemes of different types: IMEX, explicit, exponential and implicit. Their
experimentally maximal stable CFL-number CFLstable is listed as well.
The results are shown in Fig. 3. First, we verify the order of convergence in Fig. 3a: All schemes achieve
the expected slopes until machine precision interrupts. The higher order exponential schemes hit that
ground already for large timestep sizes (≈ 3e-03), since the subsequent point with the same slope would
be beyond machine precision. Moreover, we see the large difference in the maximal stable time-step sizes
for the different types of schemes: implicit or exponential, IMEX and explicit. Even though the maximal
stable CFL-number for the IMEX-schemes is a bit below the one of the explicit schemes in that testcase
305
a) Convergence study b) Efficiency study
Figure 3. Results of the computations of the Nozzle flow test case until t=5. For orientation the graphs of the
IMEX schemes are in blue, of the explicit schemes in black, of the exponential schemes in red and of
the implicit schemes in green.
(see Tab. 1), the maximal stable timestep size can be chosen 50 times larger than for explicit schemes,
since the smallest element in the CFL-condition is much larger for IMEX-schemes.
The work-precision diagram confirms the efficiency advantage of the IMEX schemes compared to non-
IMEX schemes in that the same precision can be achieved at a similar or lower cost. The novel EXPEX
schemes beat the IMEX-RK schemes with factors of up to 2.5 for the 3rd order schemes and a factor
of up to 6 for the 1st order schemes. Furthermore, the high efficiency of the exponential integrators is
demonstrated impressively, verifying that they are the key to efficiency in the EXPEX schemes, while
implicit RK schemes are the most inefficient in our setting.
Acknowledgments
We thank the German Research Foundation DFG for its financial support within the project GZ: ME
1889/7-1, by which we could employ the student assistant Georgios Tsompanidis, who did the testing and
we thank Mayya Tokman for providing online the code package EPIC with the exponential integrators.
References
[1] P. Birken. Numerical Methods for the Unsteady Compressible Navier-Stokes Equations. Habilitation
Thesis. University of Kassel, 2012.
[2] Alex Kanevsky, Mark H. Carpenter, David Gottlieb, and Jan S. Hesthaven. Application of implicit-
explicit high order Runge-Kutta methods to discontinuous-Galerkin schemes. J. Comput. Phys. 225.2
(2007).
[3] Christopher A. Kennedy and Mark H. Carpenter. Additive Runge-Kutta Schemes for Convection-
Diffusion-Reaction Equations. Applied Numerical Mathematics 44.1-2 (2003).
[4] J. Loffeld and M. Tokman. Comparative performance of exponential, implicit, and explicit integrators
for stiff systems of ODEs. J. Comput. Appl. Math. 241 (2013).
[5] Veronika Straub, Sigrun Ortleb, Philipp Birken, and Andreas Meister. Adopting (s)EPIRK schemes
in a domain-based IMEX setting. AIP Conference Proceedings 1863 (2017).
[6] Veronika Straub, Sigrun Ortleb, Philipp Birken, and Andreas Meister. On Stability and Conservation
Properties of (s)EPIRK Integrators in the Context of Discretized PDEs. In: Theory, Numerics and
Applications of Hyperbolic Problems II. Ed. by Christian Klingenberg and Michael Westdickenberg.
Vol. 237. Springer International Publishing, 2018.
[7] M. Tokman. A new class of exponential propagation iterative methods of Runge-Kutta type (EPIRK).
J. Comput. Phys. 230.24 (2011).
306

High order IMEX splitting methods for

low Mach number flows
∗ 1
Jonas Zeifang
1 IAG, University of Stuttgart, Pfaffenwaldring 21, D-70569 Stuttgart
Abstract. Flows at low Mach numbers require the use of specially designed numerical schemes. In this talk
two approaches for the efficient treatment of low Mach number flows for different applications will be pre-
sented. First, the so-called RS-IMEX splitting for the Euler equations, introduced in [2], is recalled. It relies
on the incompressible limit equations and requires a specific modification of the Euler equations to prevent the
eigenvalues of the explicit system from becoming complex. This splitting is combined with a high order discon-
tinuous Galerkin method in space and an IMEX-Runge-Kutta method in time. For solving the arising linear
system, a matrix-free GMRES method is equipped. Several testcases will show the influence of the modification
and the performance of the developed method. Second, approaches on how to enable matrix-free fully implicit
simulations of low Mach number flows in a multiphase setting and how to increase their efficiency are addressed.
Introduction
Considering the Euler equations of fluid dynamics different types of characteristic waves are present - fast
pressure waves and convection. For low Mach number flows, the ratio of the convective velocity and the
velocity of the pressure waves is very low, leading to a stiff system. Especially, when simulating multiphase
flows often situations occur where the flow field is characterized by low Mach numbers. Frequently used
explicit time integration schemes require the temporal resolution of all waves necessitating very small
timesteps as the CFL condition
∆x
∆t = CFL · , CFL < 1,
max|λ|
has to be fulfilled. For the one dimensional Euler equations the eigenvalues λ are given by
λ1,3 = u ± c, λ2 = u,
Hence, they are computationally inefficient for the simulation of low Mach number flows where u c.
Implicit schemes have the advantage of allowing much larger timesteps but require the solution of large
non-linear systems and introduce very much numerical viscosity. IMEX schemes combine both approaches
enabling an efficient treatment of low Mach number flows.
In [2] we have introduced the so-called RS-IMEX splitting for the Euler equations which allows the split-
ting of the equations in a Mach number dependent and independent part. It offers one possibility of
encountering the problems arising from the stiff system.
When simulating multiphase flows, solving the equation system becomes numerically even more challeng-
ing as very large density rations are present and more complex equations of state have to be used. Here
we will present the influence of the design of the matrix-free method on the efficiency of the scheme.
∗ Corresponding author: jonas.zeifang@iag.uni-stuttgart.de
307
1 RS-IMEX splitting for the Euler equations
The RS-IMEX splitting for the Euler equations has been introduced in [2] and will be recalled shortly.
We consider the non-dimensional Euler equations
   
ρ ρu
∂t ρu +∇ · ρu ⊗ u + εp2 Id = 0,
E u (E + p)
| {z } | {z }
=:w =:f (w)
where ρ denotes density, u velocity, E energy density and the pressure is computed via the equation of
e and f
state for a perfect gas. The RS-IMEX splitting defines the fluxes f b as
e (w) := f (wref ) + ∇w f (wref )(w − wref )

f and b (w) := f (w) − f
f e (w).
for a given reference solution wref . The choice of this splitting would result in potentially complex
b . Hence, f
eigenvalues of the Jacobian of f b is redefined as
 
0
b (w) = f (w) − f
f e (w) −  − γ−1 2
2 ρkuref − uk · I
,
ε2 (γ−1) 2 2 2

2 ρ (kuref k2 − kuk2 )(u − uref ) − ku − uref k2 uref
violating fb+f e = f , and introducing a modeling error to with respect to the Euler equations. If now
uref is chosen as the incompressible velocity field for the same initial conditions, the size of the modeling
error lies in O(ε2 ). The eigenvalues of the flux Jacobians are now real and for f b even independent of
b
the reference Mach number ε. Hence, an IMEX time discretization treating f explicitly and f e implicitly
allows the choice of the time step independent on the Mach number. In this talk we will present results
of combining the RS-IMEX splitting with a high order IMEX Runge-Kutta scheme in time and a high
order discontinuous Galerkin method in space.
2 A Jacobian-free method for multiphase simulations
Simulating flows where liquids are concerned necessitates the use of non ideal gas equation of states. The
simplest variant is the stiffened gas equation of state

1 2
p = (γ − 1) E − ρkuk2 − ρq − γp∞ ,
2
where q and p∞ are material dependent parameters. Other options such as the use of tabulated equation
of states require the use of dimensional quantities for the simulation. This results in very large absolute
values e.g. for the energy density, leading to problems with machine accuracy especially when using
implicit time discretization with a Jacobian-free approach. The implicit time discretization relies on the
solution of a non-linear equation system. This can be solved with a Newton-GMRES method based on
the approximate evaluation of the product of the Jacobian of the spatial discretization with the solution
vector of Newton’s method ∆w
e
∂ R(w) e
R(w e
+ ∆F D ∆w) − R(w)
∆w ≈ , (1)
∂w ∆F D
e denoting the spatial operator of the implicitly treated system and ∆F D being a design parameter.
with R
The choice of ∆F D has to be a trade off between truncation error of the finite difference (Eq. (1)) and
round off errors caused by limited machine accuracy. Here, we will present the influence of the choice of
∆F D which can exemplary be found in [1] and cast into the form
√
εmachine
∆F D = εscale ,
k∆wk2
308
on the accuracy of the Jacobian vector product (Eq. (1)). Another approach to increase the accuracy of
the matrix-free method is to decrease the truncation error using a second order finite difference
e
∂ R(w) e
R(w e
+ ∆F D O2 ∆w) − R(w − ∆F D O2 ∆w)
∆w ≈ .
∂w 2∆F D O2
In this talk it will be shown that for realistic multiphase settings the use of a second order finite difference
is necessary. Moreover the choice of ∆F D has a significant influence on the efficiency of the scheme.
Acknowledgments
This work has been supported by the German Research Foundation (DFG) through the International
Research Training Group GRK 2160/1: Droplet Interaction Technologies. The support and computing
time provided by the High Performance Computing Center Stuttgart (HLRS) through the hpcdg project
is acknowledged.
References
[1] D. A. Knoll and D. E. Keyes. Jacobian-free Newton–Krylov methods: A survey of approaches and
applications. Journal of Computational Physics 193 (2004), pp. 357–397.
[2] J. Zeifang, J. Schütz, K. Kaiser, A. Beck, M. Lukáčová-Medvid’ová, and S. Noelle. A novel full-Euler
low Mach number IMEX splitting. Accepted by Communications in Computational Physics (2019).
309
310
MS 17: Coupled Problems

Computation of the Stability Limit of Nonlinear Bladerow

Flutter with a Fully Coupled Frequency Domain FSI
Solver
1 2
Christian Berthold , Johann Gross , Christian Frey
1 and Malte Krack
2
1 Institute of Propulsion Technology, German Aerospace Center, Linder Höhe, 51147 Cologne, Germany , christian.berthold@dlr.de
2 Institute of Aircraft Propulsion Systems, University of Stuttgart, Pfaenwaldring 6, 70569 Stuttgart, Germany
Abstract. To guarantee a safe operation of turbomachinery components engineers need to consider the aeroe-
lastic behavior of the mechanical design. Due to increasing demands with respect to eciency and fuel
consumption components need to be further optimized. However conventional computational analysis meth-
ods for aeroelasticity which are commonly used in industry during the design phase have uncertainties since
they are often based on linearization or neglect interactions between physical nonlinear eects in the solid
and uid domain. Thus large safety margins for conventional methods need to be considered resulting in
restrictive design guidelines. Furthermore the nonlinear behavior of blade vibrations and the interaction with
the surrounding nonlinear ow is only rarely analyzed and therefore not yet fully understood. Time accurate
methods can deliver higher accuracy but the computational costs are prohibitive for the industrial design pro-
cess. In this study a fully coupled uid-structure interaction (FSI) solver in the frequency domain is presented
to provide a computationally ecient simulation tool which is able to consider nonlinear solid and uid eects
of blade limit cycle oscillations (LCOs) in turbo engines. It can compute stable LCOs as well as unstable
LCOs (stability limit) which can occur e.g. due to dry nonlinear friction in shrouded blades and is based on
the Harmonic Balance method. The fourier components of the blade vibration and the frequency are passed
to the ow solver and the aerodynamic forces are returned to the strucure solver. Because the frequency
of self excited vibrations is not known a priori the FSI solver considers the frequency as part of the sought
solution. Furthermore a linearization of the aerodynamic forces inside the structure solver is introduced to
improve the convergence of the coupled solver and countermeasures are implemented to prevent the solver from
converging to the trivial solution (zero amplitude). To further reduce the computational costs the structural
model is reduced with the Craig-Bampton approach. During the reduction the degrees of freedom in the solid
contact regions of the bladerow are retained to model dry friction. To initialize the solver the energy method
is extended to approximately determine the nonlinear behavior with a one-way coupling approach. The novel
method is applied to a bladerow which is aerodynamically stable for small vibration amplitudes. With in-
creasing amplitude the modeshape as well as the frequency change due to nonlinear friction in the joints of
the shrouded rotor. This has an eect on the aerodynamic response and for large amplitudes the oscillations
are aerodynamically unstable. The coupled solver is able to determine the limit of stability and the frequency
of the coupled problem while conventional energy methods are not able to detect a stability boundary at all.
Introduction
An important topic within aeroelasticity are self-excited vibrations of turbomachinery blades leading to
limit cycle oscillations (LCOs) [3]. The conventional design approach is to discard all designs which
exhibit utter. However, as uttering blades often vibrate at small amplitudes, this approach may be
too restrictive. On the other hand an aerodynamically stable bladerow could become unstable for large
amplitudes due to nonlinear eects which also cannot be predicted with conventional design methods
[7] since the conventional energy method is based on the linearization of the coupled system for small
vibrations.
In structural dynamics, nonlinear eects can occur due to nonlinear physical phenomena in mechanical
joints located e.g. in the shroud, blade root or by the use of underplatform dampers [5]. With increasing
amplitudes the contact conditions may change from fully sticking to a sliding motion which in turn result
in a change of shape and frequency of the structural eigenmode. On the aerodynamic side it is well-known
that the aerodynamic response depends strongly on the vibrational frequency and on the modeshape [3].
Available time integration simulation methods are computationally to expensive [2]. Frequency domain
methods such as the harmonic balance (HB) method are popular in both physical domains due to low
311
simulation costs. Current frequency domain approaches are usually one-way coupled methods. It is
therefore desirable to have a computationally ecient two-way coupled FSI solver for the analysis of
utter induced LCOs. In this study a method which enables a two-way coupled analysis of LCOs due to
utter with the use of the HB method is presented. The rened energy method [3] is used to validate the
coupled solver and also to supply a good initialization to the coupled FSI solver. In the following a hat
indicates complex Fourier coecents.
1 Benchmark model and single-domain solvers
The low pressure turbine testcase (see Fig. 1) is kindly provided by MTU Aero Engines AG. The contact
areas are modeled with nodal contact elements and are located in the shroud. For details on this modeling
approach, see [5]. The full bladed disk is assumed as rotationally periodic system. To increase the
contact area
sensor
node
cyclic
boundary
fixed
boundary
Figure 1. Structural FE model of the bladed disk [3]
computational eciency the full FEM model which models the dynamic force equilibrium is reduced
with the Craig-Bampton substructuring technique [4]. The reduced order model (ROM) is transformed
into the frequency domain with the HB technique [5]. The solution n
(i.e. the displacements of the structure
PK ikωt s o
u (t)) is approximated with a truncated fourier series u (t) ≈ <
s s
k=0 e ûk which is solved for a
a
given harmonic aerodynamic force f̂ (ûf , ω). The conservative ow variables uf = [ρ, ρu, ρv, ρw, ρE]T
(mass, x-, y -, z -momentum and energy) of the URANS equations are solved together with a k -ω turbulence
model. After applying the FV method to each cell the ow equations are subsequently transformed into
the frequency domain [1] analogous to the structural model. Both systems of equations denoted by
a
Rf (ûf , ω, χ̂) = 0 (solved with TRACE) and Rs (ûs , ω, f̂ ) = 0 (solved with NLvib) are coupled through
a
the aerodynamic force f̂ (ûf , ω) and the FV mesh motion χ̂(ûs ) due to the motion of the structure.
2 Coupled solver
The coupled FSI solver [3] consists of the uid and structure HB solvers and exchanges the Fourier com-
ponents of the solution (see Fig. 2). The vibrational frequency is considered as a part of the solution
since for self excited vibrations the exact frequency is usually not known a priori. The structural solver
has the tendency to converge to the trivial solution (static equilibrium). As countermeasure the funda-
mental harmonic ûs1 is prevented from converging to zero by setting the real part of a selected generalized
coordinate ûsj = eTj ûs1 constant as depicted in Fig. 3 (see [3] for details). The imaginary part is still
part of the solution and thus there is an implicit phase constraint which would otherwise be arbitrary
in the autonomous case. Since the coupled system is self-excited it is of autonomous nature and the
coupled frequency is part of the solution. Furthermore only by considering the frequency as unknown,
the structural algebraic system of equations is not overdetermined. The phase constraint applied to the
structural solver in conjunction with passing the updated frequency to the ow solver will enable the FSI
solver to determine the frequency during the coupling iterations.
a
Within the structural solver the aerodynamic force f̂ can be linearized (which corresponds to the con-
ventional energy method) to speed up the overall convergence of the FSI solver. This is motivated by
obeservations that show, that often the aerodynamic damping coecient does not depend signicantly
on the vibration amplitude [6]. For details on the linearization see [3].
312
structure
s
iterate
on ûa ,
ω until
s s
R û , ω, f̂ (i) < s

ω(i) − ω(i−1) < ω
yes
s s
û(i) − û(i−1) < solved
u
no
aerodynamic
force
ow mesh initialization
a
f̂ (i) = f̂
a
ûf(i−1) , ω(i−1) χ̂(i) = [χ1 χ2 . . . χN ]ûs(i) ûs(0) , ω(0)
i ← i + 1
ow
f
iterate on û until
f f
R û , ω(i−1) , χ̂(i−1) < f
Figure 2. The algorithm scheme for the FSI frequency domain solver. [3]
Im
ûsi
Re
Figure 3. Avoidance of trivial solution.
3 Testcase Results
The testcase conguration is aerodynamically stable for small vibrational amplitudes and unstable for
larger amplitudes due to a change of modeshape and eigenfrequency. Thus the coupled FSI solver is
used to determine the stability limit amplitude. The ND −4 is selected for the investigation of the
coupled nonlinear vibrational behavior. One harmonic in the structure model and the fundamental and
the second harmonic in the uid model are considered. The convergence of the rst four iterations of
the coupled FSI solver are shown in Fig. 4(a) and the convergence of ûsj is shown in Fig. 4(b). The
chosen linearization approach for the aerodynamic force performed in a robust way for this testcase
conguration. The vibrational frequency of the rened energy method (≈418 Hz) and the FSI solver
solution are signicantly dierent from the in-vacuum frequency of the linearized eigenmode (519.661 Hz).
The results of the rened energy method and the FSI solver agree very well and the deviations of frequency
and ûsj are 0.149 % and 2.137 % respectively. The mode shapes of the FSI solver solution and the rened
energy method are very similar. The coupled solver requires 83.5 hours on 10 cores.
4 Conclusions
The good agreements between the solutions of the novel FSI solver and the rened energy method verify
the implementation of the coupled solver and show its value for future investigations. The testcase con-
guration demonstrated the occurence of nonlinear instability of the coupled problem due to nonlinear
friction in the structure. Only the novel frequency domain methods are capable of detecting the insta-
bility and also predict the stability boundarys while the conventional energy method predicts a stable
313
coupled FSI solver iteration

i=1 i=2 i=3 i=4
}
}
}
}
10 -3 10 10
10 -4
5.7
418.6
10 -4 5.6
10 0 418.5
5.5 418.4
10 -5
5.4 418.3
10 -10 418.2
5.3
10 -6 418.1
5.2 418
10 -20 0 20 40 60
10 -7 coupled FSI solver iteration
(b)
10 -8 10 -30
1 1661 1 4 1 407 1 3 1 403 1 5 1 460 1 4
Pseudo time step, Newton solver step
(a)
Figure 4. ND -4: (a) Convergence of the FSI solver for the unstable LCO; (b) Convergence of the dominant DOF
for the unstable LCO. [3]
vibration. Further improvement of the coupled frequency domain FSI solver needs to be done regarding
the initialization to circumvent the use of the rened energy method completely.
Acknowledgments
The research project (FVV no. 6001308, DFG no. FR 3970/1-1) was performed by the Institute of
Propulsion Technology (ILA) at the German Aerospace Center and by the Institute of Aircraft Propulsion
Systems (ILA) at University of Stuttgart. The authors greatfully acknowledge the support received
from the FVV (Research Association for Combustion Engines eV) and the DFG (German Research
Foundation).
References
[1] G. Ashcroft, C. Frey, and H.-P. Kersken, eds. On the development of a harmonic balance method for
aeroelastic analysis.
2014.
[2] Christian Berthold, Christian Frey, Guido Dhondt, and Harald Schönenborn. FULLY COUPLED
AEROELASTIC SIMULATIONS OF LIMIT CYCLE OSCILLATIONS IN THE TIME DOMAIN.
In: Proceedings of the 15th ISUAAAT. 2018. url: https://elib.dlr.de/127137/.
[3] Christian Berthold, Johan Gross, Malte Krack, and Christian Frey. Computation of the Stability
Limit of Nonlinear Bladerow Flutter with a Fully Coupled Frequency Domain FSI Solver. Journal
of Fluids and Structures (Status: Submitted) (2019).
[4] R. R. Craig and M. C.C. Bampton. Coupling of Substructures for Dynamic Analysis. AIAA Journal
6.7 (1968), pp. 13131319.
[5] Malte Krack, Loic Salles, and Fabrice Thouverez. Vibration prediction of bladed disks coupled by
friction joints. Archives of Computational Methods in Engineering 24.3 (2017), pp. 589636.
[6] Charles E. Seeley, Christian Wakelam, Xuefeng Zhang, Douglas Hofer, and Wei-Min Ren. Investi-
gations of Flutter and Aero Damping of a Turbine Blade: Part 1 - Experimental Characterization.
In: ASME Turbo Expo 2016: Turbomachinery Technical Conference and Exposition. 49842. 2016,
V07BT34A024. doi: 10.1115/GT2016-57930. url: http://dx.doi.org/10.1115/GT2016-57930.
[7] X Wu, M Vahdati, C Schipani, and M Imregun. Analysis of Low-Pressure Turbine Flutter for Dif-
ferent Shroud Interfaces. In: ASME Turbo Expo 2007: Power for Land, Sea, and Air. American
Society of Mechanical Engineers. 2007, pp. 629636.
314

Adaptive Multilevel Optimization of Fluid-Structure In-

teraction Problems
Johanna K. Biehl1 and Stefan Ulbrich1
1 FB Mathematik, AG Optimierung, TU Darmstadt
Abstract. We present an effective way to control fluid-structure interaction problems in a two dimensional
setting, where the fluid is modeled with the Navier–Stokes equations and the structure is hyperelastic. To
this end, we first introduce the model and the coupling conditions used. Then, a weak formulation is derived,
which is the basis for a monolithic finite element discretization in space and a Crank-Nicolson scheme in time.
Since the discretization results in a problem which is costly and time consuming to solve, for an effective
optimization, we want to reduce the computational time. Therefore, we introduce a reduced order model
based on proper orthogonal decomposition. The stability of solutions of the resulting reduced saddle point
problem is ensured with help of supremizers, which guarantee that the inf-sup condition is satisfied. For the
optimization we then introduce an adjoint based adaptive multilevel SQP method. Here, the accuracy of the
finite element model is further ensured by adaptive grid refinement strategies which result from residual based
a posteriori error estimators.
Introduction
The interaction between fluids and structures is a widespread problem in many applications, as for
example in medicine and civil engineering. We present a multilevel SQP method to control these kind of
problems. We consider a two dimensional setting, where the fluid is modeled as a Newtonian fluid by the
Navier–Stokes equations and the structure is hyperelastic, e.g., a Saint Venant–Kirchhoff material. We
first introduce the model from which we derive a weak formulation. This is the basis for a monolithic
finite element discretization in space and a Crank-Nicolson scheme in time. This discretization results in
a high dimensional, highly nonlinear problem. For an effective optimization, we therefore want to reduce
the computational cost. This is done by introducing a reduced order model based on the discrete proper
orthogonal decomposition technique. The resulting system is a saddle point problem and is extended by
supremizers to ensure the stability of solutions by guaranteeing that the inf-sup condition is satisfied.
For the optimization an adjoint based adaptive multilevel sequential quadratic programming method
is used. The basic idea is to compute the optimization steps on the reduced order model and control
their accuracy on the finite element level. We further ensure the accuracy of the finite element model
by adaptive grid refinement strategies, which result from residual based a posteriori error estimators.
Finally, the presented method is applied to the Benchmark problem proposed by Turek and Hron in [7].
1 Problem modelling
The fluid is considered to be Newtonian and, as already mentioned, modeled by the Navier–Stokes
equations, consisting of mass and momentum conservation equations. These equations depend on the
pressure p and the velocity v of the flow. This flow interacts with the hyperelastic structure, given by the
momentum conservation, e.g., the nonlinear elasticity equation, which is solved for the deformation u. To
overcome the difficulty that usually fluids are modeled in a Eulerian and solids in a Lagrangian setting,
an Arbitrary Lagrangian Eulerian (ALE) formulation is used. This can be done in two ways, either all
terms are transformed onto the reference domain or the fluid is modeled on the deformed domain. Here,
we concentrate on the first option and denote with a hat symbol whenever a function or variable is defined
on the reference configuration.
315
As interaction conditions on the interface we impose the continuity of velocities ∂t u = v and the action
equals reaction condition for the normal stresses. We further choose a suitable do-nothing boundary
condition for the flow on the outflow boundary and no-slip conditions, a Dirichlet inflow condition and
no-slip conditions on the remaining boundaries.
For the boundary of the solid which is not part of the interface we impose homogeneous Dirichlet boundary
conditions. All together this gives the following initial value problem for a time interval T := [0, tend ].
Problem 1 Find û : Ω̂s × T → R2 , v̂ : Ω̂f × T → R2 and p : Ω̂f × T → R, such that for suitable initial
data û(x̂, 0) = û0 (x̂) in Ω̂s and v̂(x̂, 0) = v̂0 (x̂) in Ω̂f it holds:
2
ρs ∂tt c (Jˆσ̂s F̂ −T ) = ρs fˆs ,
û − div
ˆ f ∂t v̂ + Jρ
Jρ ˆ f (F̂ −1 (v̂ − ∂t T̂ ) · ∇)v̂ c (Jˆσ̂f F̂ −T ) = Jρ
b − div ˆ f fˆf ,
c (JˆF̂ −1 v̂) = 0,
div
with constant densities ρs , ρf , interaction conditions
Jˆσ̂s F̂ −T n̂s = Jˆσ̂f F̂ −T n̂f , on Γ̂I × T,

∂t û = v̂, on Γ̂I × T,
and boundary conditions
v̂ = v̂ in on Γ̂D
in × T,
v̂ = 0 on Γ̂D
f × T,
ˆ F̂ · n̂ − p̂Iˆ = 0
ρf νf ∇v̂ −1
on Γ̂out
N × T,
û = 0 on Γ̂D
s × T.
The stress tensors are given by
σ̂s = Jˆ−1 F̂ (λ tr (Ê)I + 2µÊ)F̂ T ,

σ̂f = −p̂Iˆ + ρf νf (∇v̂
ˆ F̂ −1 + F̂ −T ∇v̂
ˆ T ).
Here, T̂ , the ALE transformation, is a diffeomorphism extending the deformation T̂L := x̂ + û(x̂, t), F̂ is
the so-called deformation gradient and Jˆ its determinant. Further, Ê denotes the Green–Lagrange strain
tensor. The densities are given by the constants ρf and ρs and the material is defined by using the Lamé
constants µ and λ.
The basis for the discretization is a weak formulation of the problem. This is derived by multiplying the
equations with suitable test functions and integrating over the domain. Then, by integration by parts,
we can shift some derivatives onto the test functions. The so derived weak system is then discretized by
a finite element approach in space and a Crank-Nicolson scheme in time, following [3].
2 Reduced order model
To reduce the costs of the computation, we use model order reduction, more specifically, the proper
orthogonal decomposition method. The basic idea of this method is to approximate the state by ap-
proximating the space spanned by a given number of time snapshots. This method has already been
successfully used for a variety of optimal control problems, see, e.g., [2, 5].
As the finite element discretization, also the reduction for fluid structure interaction problems results
in a saddle point problem. Existence and stability of a solution for such problems can be ensured with
Brezzi’s Splitting Theorem. This means we have to make sure that the inf-sup condition holds true for
the derived reduced model. This does not need to be the case for the standard choice of the POD basis.
Thus, we follow the proposal of Rozza and Veroy in [6] and enrich the POD basis.
Since in general we are looking at optimization problems and calculate the derivative of the cost function
by using adjoint calculation, we also reduce the space of the adjoint states.
316
3 Optimization
For the optimization we use a multilevel sequential quadratic programming method, with adaptive mesh
refinement. The method is based on the works of Ziems, Ulbrich and Bott in [8, 9] and [1].
Therefore, we need to control the finite element error as well as the error occuring through the model order
reduction. For the finite element error, we derive an estimator which is close to the one in [4] for the time
independent Navier–Stokes equations. We further control the error introduced by the time discretization
by a suitable routine. The idea of the second error estimator is to evaluate the finite element system with
the solutions of the reduced order model. This gives reasonable estimates, see [1].
4 Application
The setting of the Benchmark problem and the notation is shown in Figure 1. The structure with domain
Ω̂s is attached to a fixed cylinder in a two dimensional channel flow Ω̂f , with interface Γ̂I , inflow boundary
in , and outflow boundary Γ̂out .
Γ̂D N
Γ̂D
f Ω̂s
Γ̂D Ω̂f Γ̂D
f Γ̂N
in out
Γ̂D
s
Γ̂I
(0, 0)
Figure 1. Benchmark problem.
The implementation of the finite element model is in good agreement to the in the benchmark proposed
values. For the test of the optimization algorithm we first have to introduce a control z and a suitable
control space Z as well as a cost functional J. For the control we choose a body force acting on a given
subdomain of the fluid Ω̂f , for which we introduce box constraints. As an example for a cost functional,
we choose the lift forces acting on the interface and cylinder.
·10−2
1,000
6
Reduced gradient
Cost function
4
500
0
0
20 40 60 80
20 40 60 80
Iteration
Iteration
Figure 2. Development of cost function and reduced gradient
In Figure 2, we see the development of the cost function and the reduced gradient over the iterations of
the multilevel SQP method. The squares always denote a finite element evaluation. All other steps are
computed on a reduced order model. The original and optimized lift forces are shown in Figure 3.
Acknowledgments
This work is supported by the ‘Excellence Initiative’ of the German Federal and State Governments and
the Graduate School of Computational Engineering at Technische Universität Darmstadt.
317
200
100
lift
0
−100
optimized
original
−200
0 0.2 0.4 0.6 0.8 1
time
Figure 3. Original and reduced lift forces
References
[1] S. Bott. “Adaptive SQP method with reduced order models for optimal control problems with con-
straints on the state applied to the Navier–Stokes Equations”. PhD thesis. Technischen Universität
Darmstadt, 2016.
[2] K. Gubisch and S. Volkwein. Proper orthogonal decomposition for linear–quadraticuadratic optimal
control (2013).
[3] Jaroslav Hron and Stefan Turek. A monolithic FEM/multigrid solver for an ALE formulation of fluid-
structure interaction with applications in biomechanics. English. In: Fluid-structure interaction.
Modelling, simulation, optimisation. Proceedings of the workshop, Hohenwart, Germany, October
2005. Berlin: Springer, 2006, pp. 146–170. isbn: 3-540-34595-7/pbk.
[4] Volker John. Residual a posteriori error estimates for two-level finite element methods for the Navier-
Stokes equations. English. Appl. Numer. Math. 37.4 (2001), pp. 503–518. issn: 0168-9274. doi:
10.1016/S0168-9274(00)00058-1.
[5] K. Kunisch and S. Volkwein. Proper orthogonal decomposition methods for a general equation in
fluid dynamics. SIAM J. Numer. Anal. 40.2 (2002), pp. 492–515. doi: 10.1137/S0036142900382612.
url: http://dx.doi.org/10.1137/S0036142900382612.
[6] Gianluigi Rozza and Karen Veroy. On the stability of the reduced basis method for Stokes equations
in parametrized domains. English. Comput. Methods Appl. Mech. Eng. 196.7 (2007), pp. 1244–1260.
issn: 0045-7825. doi: 10.1016/j.cma.2006.09.005.
[7] Stefan Turek and Jaroslav Hron. Proposal for numerical benchmarking of fluid-structure interaction
between an elastic object and laminar incompressible flow. English. In: Fluid-Structure Interaction.
Vol. 53. Lecture Notes in Computational Science and Engineering. Springer Berlin Heidelberg, 2006,
pp. 371–385. isbn: 978-3-540-34595-4. doi: 10.1007/3- 540- 34596- 5_15. url: http://dx.doi.
org/10.1007/3-540-34596-5_15.
[8] Stefan Ulbrich and Jan Carsten Ziems. Adaptive multilevel trust-region methods for time-dependent
PDE-constrained optimization. English. Port. Math. (N.S.) 74.1 (2017), pp. 37–67. issn: 0032-5155;
1662-2758/e.
[9] C. Ziems. “Adaptive multilevel SQP–methods for PDE–constrained optimization”. PhD thesis. Tech-
nischen Universität Darmstadt, 2010.
318

Finite element modelling of ultrasonic guided waves

propagation in the fluid-composite structure interaction
1
Bhuiyan Shameem Mahmood Ebna Hai ∗
and Markus Bause1
1 Numerical Mathematics, Faculty of Mechanical Engineering, Helmut Schmidt University, Hamburg, Germany.
Abstract. This work proposes a concept of coupling the ultrasonic guided waves (UGWs) propagation
with the fluid-composite structure interaction, where the composite elastic structure consists of at least
two layers. This coupled multiphysics model tracks the UGWs propagation through the fluid, compos-
ite/multilayer structure, and their interface as a part of the structural health monitoring (SHM) system.
The model includes a monolithically coupled system of acoustic and elastic wave equations. The solid part
comprises the Saint-Venant Kirchhoff (STVK) and neo-Hookean (INH) material model. The coupling be-
tween the two solid layers as well as between the multilayer structure and fluid inclusion is achieved via
the geometric, kinematic, dynamic, interface regularity and boundary conditions. To ensure that the mul-
tilayer structure and fluid domains are conforming, we use the Arbitrary Lagrangian-Eulerian (ALE) tech-
nique. The numerical solution approach is based on the combination of Finite Element and Finite Dif-
ference methods. We follow the Rothe method, in which discretization in time is followed by space dis-
cretization, solving the resulting problem according to the monolithic approach. Discretization in space
is done via a Finite Element method. An application example discussed here offers computational sup-
port for the off-line (or passive) SHM systems for a composite/multilayer material and lightweight structure.
Rationale
# $%&&
𝜕Ω
"#
Ω
"%
𝜕Ω "%
𝜕Ω
#'
Ω
"&
Ω "&
Ω
' (
# *+,
𝜕Ω
# ()
𝜕Ω #.
Ω #.
Ω
3 3 "%
𝜕Ω "&
𝜕Ω "%
𝜕Ω
,
#.
Ω 𝑓*+&
4
#(
𝜕Ω
𝑓1$.
"%
𝜕Ω "%
𝜕Ω
#.
𝜕Ω 2
# $%&&
𝜕Ω # $%&&
𝜕Ω
# -%./
𝜕Ω
(a) (b)
Figure 1. Configuration of 2D solid plate with a disc-shaped burst signal fˆws : (a) a elastic body with three layers
under fluid flow, and (b) a composite solid plate with two butt-joint solids.
While structural damage can be a continuously process that spreads during the operation of the object,
a complete failure is usually a sudden event. As history shows, the consequences can be catastrophic.
Therefore, it is of utmost importance to predict and ultimately avoid the occurrence of such events. It
is precisely the motivation behind the design of an online SHM system. In addition, ideally, concerns
for safety should not put a properly functioning structure out-of-service. Thereby, online SHM system is
also a matter of cost-effectiveness. Furthermore, the online SHM system should be able to facilitate an
operation during the unexpected or expected external forcing. For example, during such extreme events
as earthquake, explosion, storm, collision, etc. an online SHM system can be used to detect, evaluate
the extent of the damage, and decide whether the use of the mechanical structure is safe. The eXtended
fluid-structure interaction (eXFSI) model constitutes a first step towards the computational support for
∗ Corresponding author: ebnahaib@hsu-hh.de ORCID iD: 0000-0002-7747-5536
319
the online SHM system and finds it’s use in a wide range of applications [1, 2, 3, 4, 5, 6, 7, 8]. This
work contributes to the emerging strand of literature focusing on computational techniques, supporting
the SHM system design.
To start with, we specify the computational domain. Let us consider, the given test configuration Ω̂
at time t = 0, t ∈ I, and T > 0, be composed of two time-dependent subdomains: solid domain
Ω̂s = Ω̂sb ∪ Ω̂sd , and fluid domain Ω̂f . Here, Ω̂ = Ω̂s ∪ Ω̂f , and Ω̂s ∩ Ω̂f = ∅ (cf. Figure 1). The
reference (or initial) domain Ω̂ has a Lipschitz boundary ∂ Ω̂ ⊂ Rd−1 , which is subdivided into four parts
as follows: ∂ Ω̂ = ∂ Ω̂in ∪ ∂ Ω̂wall ∪ ∂ Ω̂base ∪ ∂ Ω̂out , where for the displacements and velocities ∂ Ω̂wall ,
∂ Ω̂in , and ∂ Ω̂base are consider as Dirichlet boundaries, and ∂ Ω̂out denotes as the fluid outflow Neumann
boundary. The common fluid-solid interface ∂ Ω̂i ⊂ Rd−1 and solid-solid interface ∂ Ω̂si ⊂ Rd−1 are given
by ∂Ωi = Ω̂f ∩ Ω̂s and ∂Ωsi = Ω̂sb ∩ Ω̂sd , respectively. Since we are using a variational monolithic
coupling method, the displacements and velocity spaces are extended to the entire domain Ω̂. For the
eXFSI problem, we work with the following spaces:
L̂Ω̂X := L2 (Ω̂X ), L̂0Ω̂X := L2 (Ω̂X )/R, V̂Ω̂X := H 1 (Ω̂X ), V̂Ω̂0X := H01 (Ω̂X ).
Here, X = {s, f } is a set of suffixes, where “f ” stands for the fluid and “s” stands for the solid. For
convenience, we omit the explicit time-dependence term when indicating the time-dependent domains
i.e. Ω := Ω(t). Last, but not least, for the current domains Ω, Ωf and Ωs , we can denote the corre-
sponding spaces without the ’hat’ (ˆ) notation. Next, we introduce the FSI problem that is fully coupled
with the WpFSI problem [9] – eXtended fluid-structure interaction (eXFSI) problem. Accordingly, we
combine the following models: (i) nonlinear fluid-structure interaction in the ALE framework, including
the incompressible Navier-Stokes in the ALE framework, elastodynamics in Lagrangian framework, and
the biharmonic mesh moving model; and (ii) nonlinear wave propagation (WpFSI) problem in the ALE
framework. The latter is a coupled problem of the time-harmonic elastic wave propagation with a signal
force in the ALE framework and acoustic wave propagation in the ALE framework [9]. The novelties of
the present work include: formulation of a coupled wave propagation problem in the ALE framework,
coupling it to the FSI model, where the solid part comprises the STVK and INH material model. Next,
the model is used to analyse the implications of FSI for the wave propagation on solid-solid interface.
The eXFSI problem
Let us consider, v̂ D and ûD are suitable extensions of Dirichlet inflow data. Find the global velocity
v̂ ∈ {v̂ D +V̂Ω̂0 }, displacement û ∈ {ûD +V̂Ω̂0 }, auxiliary variable ξˆ ∈ V̂Ω̂ , pressure p̂ ∈ L̂0Ω̂ , wave displacement
ûw ∈ V̂Ω̂0 , wave velocity v̂w ∈ L̂Ω̂ , and wave pressure p̂w ∈ L̂0Ω̂ , such that the initial conditions v̂(0) = v̂ 0 ,
û(0) = û0 , ûw (0) = û0w , and ∂t ûw (0) = v̂w
0
are satisfied, and for almost all time t ∈ I it holds that:


 (Jˆρ̂f ∂t v̂, φ̂v )Ω̂f − hJˆĝf F −T , φ̂v i∂ Ω̂N



 +(Jˆρ̂f (F̂ −1 (v̂ − w) · ∇)v̂, ˆ φ̂v )Ω̂f






 +(Jˆσ̂f F̂ −T , ∇ ˆ φ̂v ) + (ρ̂s ∂t v̂, φ̂v )
Ω̂f Ω̂s


+(Jˆσ̂s F̂ , ∇
 −T ˆ φ̂v ) =0 ∀φ̂v ∈ V̂Ω̂0 ,
 Ω̂s
FSI part (αu ξ, ˆ φ̂ξ ) − (αu ∇û, ˆ ∇ ˆ φ̂ξ )
Ω̂f Ω̂f



 +(α ˆ
ξ, φ̂ ξ
) − (α ˆ
∇û, ˆ φ̂ξ )
∇ =0 ∀φ̂ξ ∈ V̂Ω̂ ,

 u Ω̂s u Ω̂s

(ρ̂s (∂t û − v̂), φ̂u ) + (αu ∇ ˆ∇
ˆ ξ, ˆ φ̂u )

 Ω̂ Ω̂f =0 ∀φ̂u ∈ V̂Ω̂0 ,


s
 ˆ JˆF̂ −1 v̂), φ̂p ) + (p̂s , φ̂p )
(div( =0 ∀φ̂p ∈ L̂0Ω̂ , (STVK)

 Ω̂f Ω̂s
 ˆ ˆ −1
 (div(J F̂ v̂), φ̂p )Ω̂f + (det F̂ − 1, φ̂p )Ω̂s =0 ∀φ̂p ∈ L̂0Ω̂ , (INH)



 (ρ̂s ∂t ûw , φ̂uw )Ω̂s − (ρ̂s v̂w , φ̂uw )Ω̂s




+(Jˆρ̂f ∂t ûw , φ̂uw )Ω̂f − (Jˆρ̂f v̂w , φ̂uw )Ω̂f




 −(Jˆρ̂f (F̂ −1 ŵ · ∇)û ˆ w , φ̂uw )
Ω̂f =0 ∀φ̂uw ∈ L̂Ω̂ ,


(ρ̂s ∂t v̂w , φ̂ w ) + (Jˆσ̂ws F̂ −T , ∇
 v ˆ φ̂vw )
 Ω̂s Ω̂s
WpFSI part −(Jˆfˆws , φ̂vw )Ω̂s + (Jˆρ̂f ∂t v̂w , φ̂vw )Ω̂f



−(Jˆρ̂f (F̂ −1 ŵ · ∇)v̂ ˆ w , φ̂vw )

 Ω̂f

+(c2 Jˆρ̂ (∇û ˆ w F̂ −1 )F̂ −T , ∇ ˆ φ̂vw )

 f =0 ∀φ̂vw ∈ V̂Ω̂0 ,


Ω̂f

(p̂ , φ̂ pw
) + (det F̂ − 1, φ̂ pw
) =0 ∀φ̂pw ∈ L̂0Ω̂ ,

 w s Ω̂s Ω̂s
|
 {z } | {z }
STVK INH
320
with the linearized stress tensor given by

ˆ f F̂ −1 + F̂ −T ∇v̂
σ̂f = −p̂f I +ρ̂f νf (∇v̂ ˆ fT ),
(
Jˆ−1 F̂ 2µs Ês + λs tr(Ês ) I F̂ T (STVK),
σ̂s =
p̂s I +µs (F̂ F̂ T − I) (INH),

1 ˆ
Ês = ∇ûs F̂ −1 + F̂ −T ∇û
ˆ Ts ,
2
(
Jˆ−1 F̂ 2µs Êws + λs tr(Êws ) I F̂ T (STVK),
σ̂ws =
p̂ws I +µs (F̂ F̂ T − I) (INH),
1 ˆ
Êws = ∇ûws F̂ −1 + F̂ −T ∇û
ˆ Tws .
2
where a deformation gradient is given by F̂ = I + ∇û ˆ s and, in turn, the deformation determinant Jˆ is a
function of F̂ . The value of ŵ is determined from the FSI problem. Furthermore, a correction term for the
ˆ T n̂f on ∂ Ω̂out . The ultimate goal of tests
’do-nothing’ outflow condition is given by ĝf = −ρ̂f νf F̂ −T ∇v̂ f
conducted in this work is to track the wave propagation in solid-solid interface. In particular, we consider
the wave propagation problem in the deformable elastic structure with two butt-joint solids, where the
solid part constitutes the STVK compressible material. In this proceedings, we consider the ideal case,
where the structure remains in its initial position and the acting force, i.e. the FSI effect, is negligible.
Thereby, there is no solid mesh deformation due to the additional force. We apply the Rothe method
for the problem discretization. The temporal discretization is performed by using the finite difference
scheme, and for the spatial discretization we refer to a standard Galerkin finite element approach. This
nonlinear problem is solved using a Newton-like method. Worth noting, we refer to the direct solver
UMFPACK [10] for solution of the linear equation systems at each Newton step. The implementation is
accomplished via the software library package DOpElib [11].
Numerical simulation
(i) t = 13.33 ms (i) t = 13.33 ms (i) t = 13.33 ms
(ii) t = 20.0 ms (ii) t = 20.0 ms (ii) t = 20.0 ms

(a) ρsb < ρsd (b) ρsb > ρsd (c) ρsb = ρsd
Figure 2. The displacement field component ûx1 in a solid plate with two butt-joint solids.
Let us consider an elastic base plate with two butt-joint solid domains (cf. Figure-1(b)). We use the non-
vanishing disc-shaped piezoelectric actuator fˆws (x̂, t) with a radius r̂fws of 7.5 mm, which is located at
the geometric centre (0, 0) of the solid plate. The actuator is used for the excitation of a burst Lamb wave
package. Here, for the numerical simulation tests, we consider a five-cycle Lamb wave package, which has
two main displacement field components ûx1 and ûx2 in the plane. The non-vanishing burst signal force
is given in [1, 2, 3, 4] where the signal frequency (fc ) is 1000 Hz, and for the numerical simulation test, a
five-cycle Lamb wave package (ns = 5) is used. It has two UGWs displacement field components ûx1 and
ûx2 in the plane. Figure-2 illustrates the displacement component ûx1 of a five cycle Lamb wave burst
signal for various relations between the density of a base solid plate and the density of a secondary solid
321
domain. Due to the different propagation rate, the UGWs passes through the secondary solid domain at
a different time and with a different wavelength [1, 2, 3, 4, 5]. In addition, as the wave propagates from
the base domain to the secondary domain through the common solid-solid interface, strong reflections
occur. We demonstrate that this result is robust with respect to alternative material specifications, which
induce differences only in terms of speed of propagation and the amount of reflections.
Remarks
In the present work, we propose a coupled multiphysics model which supports the design of the SHM sys-
tem. We demonstrate the alterations to the wave propagation pattern for alternative media. Specifically,
we mimic the alternative configuration of a butt-joint solid plates to explore the wave propagation be-
haviour on the solid-solid interface. We show that our proposed model successfully depicts the differences
in the reflectivity between two solids and thereby is in agreement with the fundamental theory. The simu-
lation results clearly show the successful identification of the location of the solid-solid common interface.
In the present work we limit our focus on the STVK solid-solid butt-joint edge. Yet, one can extend
the numerical simulation results to identification of the structural damage, for which the fluid inclusion
can be considered [1, 3, 4]. More complete simulation results will be available in the extended version
of this proceeding. Noteworthy, the WpFSI model serves as a first step towards the implementation of a
machine learning approach to SHM system design.
References
[1] Ebna Hai, B. S. M. 2017. “Finite Element Approximation of Ultrasonic Wave Propagation under Fluid-
Structure Interaction for Structural Health Monitoring Systems", Doctoral dissertation, Helmut Schmidt
University, Hamburg, Germany.
[2] Ebna Hai, B. S. M., and M. Bause. 2019. “Numerical study and comparison of alternative time discretiza-
tion schemes for an ultrasonic guided wave propagation problem coupled with fluid–structure interaction".
Comput Math Appl.
[3] Ebna Hai, B. S. M., and M. Bause. 2019. “Modeling Concept and Numerical Simulation of Ultrasonic Waves
Propagation in a Moving Fluid-Structure Domain based on a Monolithic Approach". Appl Math Model.
[4] Ebna Hai, B. S. M., and M. Bause. 2018. “Finite Element Approximation of Fluid-Structure Interaction
(FSI) Problem with Coupled Wave Propagation". PAMM, vol. 17, pp. 511–512.
[5] Ebna Hai, B. S. M., and M. Bause. 2017. “Mathematical Modelling and Numerical Simulation of the online
Structural Health Monitoring System". In F.-K. Chang and F. Kopsaftopoulos (eds.), Structural Health
Monitoring 2017, vol. 1, pp. 580–587.
[6] Ebna Hai, B. S. M., and M. Bause. 2018. “Numerical Modeling and Approximation of the Coupling Lamb
Wave Propagation With Fluid-Structure Interaction Problem". In proceedings of: the ASME 2018 Interna-
tional Mechanical Engineering Congress & Exposition, Pittsburgh, USA, vol. 9.
[7] Ebna Hai, B. S. M., M. Bause, and P. A. Kuberry. 2016. “Finite Element Approximation of the eXtended
Fluid-Structure Interaction (eXFSI) Problem". In proceedings of: the ASME 2016 Fluids Engineering Divi-
sion Summer Meeting, Washington, D.C., USA, vol. 1A.
[8] Ebna Hai, B. S. M., and M. Bause. 2018. “Mathematical Modelling of Structural Health Monitoring Systems:
Coupling Fluid-Structure Interaction with Wave Propagation". In proceedings of: the ECCOMAS ECCM-
ECFD 2018, Glasgow, UK, pp. 394–404.
[9] Ebna Hai, B. S. M., and M. Bause. 2019. “Modeling and Simulation of Ultrasonic Guided Waves Propagation
in the Fluid-Structure Domain by a Monolithic Approach". J Fluid Struct, vol. 88, pp. 100-121.
[10] Davis, T. A. and I. S. Duff. 2007. “An unsymmetric-pattern multifrontal method for sparse LU factorization".
SIAM Journal on Matrix Analysis and Applications, vol. 18, pp. 140–158.
[11] C. Goll, T. Wick. and W. Wollner. 2017. “DOpElib: Differential Equations and Optimization
Environment". Archive of Numerical Software, vol. 5, pp. 1–14.
322

An ecient implementation of gradient-enhanced crys-

tal plasticity models
∗
V. Fohrmeister and J. Mosler
Institute of Mechanics, TU Dortmund University, Germany
Abstract. For the physically sound simulation of forming processes and a prediction of evolving textures,
micromechanically motivated crystal-plasticity models are frequently applied. If these models are enhanced
with gradients of the plastic deformations, they can capture hardening due to geometrically necessary dis-
locations. This is, for instance, important for the consideration of the Hall-Petch relation. However and in
line with the underlying local crystal plasticity model, a straightforward implementation of rate-independent
models leads to a complex and potentially ill-posed system of inequalities at the global nite elment level
(see, e.g., [6]). In order to eliminate this problem, the inequalities corresponding to the loading/unloading
conditions are rewritten by means of equivalent nonlinear complementarity conditions (NCP) equations, cf.
[8]. Furthermore and in order to reduce the implementational eort, the micromorphic approach [1] is adapted.
1 Local Crystal-Plasticity Variational formulation
1.1 Fundamentals
The framework is based on the multiplicative split of the deformation gradient
F = 1 + ∇u = F e · F p (1)
into incompatible elastic and plastic components F e and F p . With free Helmholtz energy Ψ = Ψ̂(F e , α),
where α are internal variables describing the hardening within each slip system, we can dene the Mandel
stress and the corresponding thermodynamic driving forces to α as
Σ := −∂F p Ψ · [F p ]T and Q := −∂α Ψ . (2)
By using these denitions and in line with the well-known Schmid-law, a yield criterion for every slip-
system (i) of type h i
(i)
φ(i) = M (i) · Σ · N (i) − Q0 + Q(i) . (3)
is used. Here, M (i) and N (i) are orthonormal vectors dening the slip direction and slip plane, respec-
(i)
tively. Q0 denotes the initial yield limit. These yield functions dene an outer bound for the space of
admissible stresses n o
E := (Σ, Q) ∈ R(3×3)+nsys | φ(i) (Σ, Q(i) ) ≤ 0 ∀i . (4)
Utilising the postulate of maximum dissipation, the evolution equations for the internal variables result
in
nsys h i
p X
Ḟ · [F p ]−1 = γ̇ (i) M (i) ⊗ N (i) with γ̇ (i) = sign Σ : M (i) ⊗ N (i) λ(i) , (5)
i=1
α̇ = −λ , (6)
where λ(i) denotes the plastic

Pnsys multiplier of the i-th slip system. The internal dissipation corresponding
to this model is Dint = i=1 Q(i) λ(i) .
∗ Corresponding author: volker.fohrmeister@tu-dortmund.de ORCID iD: 0000-0001-8467-9863
323
1.2 Variational formulation
For the isothermal setting, and in line with e.g. [7], rate potential
Z h i
İ := Ψ̇ + Dint dV + Pext (7)
Ω
is introduced. Here, Pext is the powers associated with external forces acting on the body Ω. The
integral term is referred to as the stress power and can be expressed as P : Ḟ for the local formulation
with P = ∂F Ψ being the rst Piola stress conjugated to F . The underlying equations of the physical
problem now follow directly by the stationary conditions with respect to the velocity u̇ and the plastic
multipliers λ, e.g.
Z
δu̇ İ = P : δ Ḟ dV + δu̇ Pext = 0 and δλ(i) İ = −φ(i) ≥ 0 . (8)
Ω
While the rst equation corresponds to the balance of linear momentum, the second one assures admis-
sibility of the stresses, see Eq. (4).
1.3 Implementation based on a nonlinear complementary problem
After applying a backward-Euler time integration

R t to the aforementioned equations, the optimality w.r.t.
the time integrated plastic multipliers ∆λ := tnn+1 λ dt can be expressed in terms of a set of Karush-
Kuhn-Tucker conditions
∂∆λ(i) I ≥ 0 ∧ ∆λ(i) ≥ 0 ∧ ∂∆λ(i) I ∆λ(i) = 0 ∀i . (9)
For numerical reasons, these inequalities are rewritten into an equivalent nonlinear complementary prob-
lem of type q
!
φ̃(i) := [∂∆λ(i) I (i) ]2 + [∆λ(i) ]2 + 2 δ 2 − ∂∆λ(i) I (i) − ∆λ(i) = 0 ∀i , (10)
i.e., the NCP functions according to Fischer and Burmeister are employed, see [4, 8]. Here, small pertur-
bation parameter δ circumvents the singularity of Eq. (10) in its origin, albeit approximating the original
problem, see[2] for further information.
2 Gradient enhancement
For the incorporation of size eects into the presented crystal plasticity model, the model is enhanced
by non-local contributions of the free energy. To be more precise, and following the idea of so-called
micromorphic models, as presented e.g. in [1], a macroscopic eld ϕp is introduced and coupled to the
internal variable α using penalty like energy Ψintr (ϕp , α). Now, gradient dependent energies Ψglob (∇ϕp )
can be introduced into the model, without the need for modifying the underlying local constitutive update
signicantly.
3 Implementation based on Hyper-Dual Step Dierentiation
For an ecient implementation of the aforementioned optimization problem, a Newton-type iteration

scheme is employed. In order to compute the required rst and second derivatives, an automatic deriva-
tion scheme based on so called hyper-dual numbers is elaborated, see [3, 5, 9]. In analogy to standard
complex numbers, this algebra contains non-real, so called dual, components j with the properties
k j = j k and 2j = 0. Analysing the Taylor-series of an arbitrary function f (x), one can show that the
rst and second derivatives of this function can be extracted by

d f D1 f (x0 + a 1 ) d2 f D12 f (x0 + a 1 + b 2 )
= , 2 = .
d x x0 a d x x0 ab
D1 (•) here extracts the coordinate in 1 direction and D12 (•) extracts the coordinate of the mixed
direction 1 2 , analogously. In contrast to other numerical schemes, the obtained derivatives are exact.
324
(b) Deformed conguration and distribution of kαk (left) and

kϕp k (right). The plots correspont to the nal state of de-
(a) Load-displacement curve of a tensile test with local formation. The three computations correspond to c = 1 Pa,
and non-local material behaviour. c = 50 Pa and c = 0.
Figure 1. Simulation of a tensile test of a single-crystal specimen.
4 Numerical Example
Simulations have been carried out with a single-crystal tensile specimen with length 96 µm and width
10 µm. The crystal orientation is not aligned with the specimen axes. For the elastic energy, a Neo-Hooke
model is used, while the coupling and penalizing energy contributions have been assumed quadratically.
Ideal plasticity is considered.
One can clearly see the performance of the non-local model in Figure 1. The local simulation stops pre-
maturely due to convergence diculties, while the non-local simulations are able to show the interaction
between geometrical softening and gradient-induced hardening.
Acknowledgement
Financial support from the German Research Foundation (DFG) via SFB/TR TRR 188 (project number
278868966), project C04, is gratefully acknowledged.
References
[1] O. Aslan, N.M. Cordero, A. Gaubert, and S. Forest. Micromorphic approach to single crystal plas-
ticity and damage. International Journal of Engineering Science 49.12 (2011), pp. 13111325. doi:
10.1016/j.ijengsci.2011.03.008.
[2] Thorsten Bartel, Robin Schulte, Andreas Menzel, Björn Kiefer, and Bob Svendsen. Investigations
on enhanced FischerBurmeister NCP functions: application to a rate-dependent model for ferro-
electrics. Archive of Applied Mechanics (Sept. 2018). doi: 10.1007/s00419-0181466-7.
[3] J. A. Fike and J. J. Alonso. Automatic Dierentiation Through the Use of Hyper-Dual Numbers for
Second Derivatives. In: Recent Advances in Algorithmic Dierentiation. Ed. by S. Forth, P. Hovland,
E. Phipps, J. Utke, A. Walther, T. J. Barth, M. Griebel, D. E. Keyes, R. M. Nieminen, D. Roose,
and T. Schlick. Vol. 87. Lecture Notes in Computational Science and Engineering. Springer Berlin
Heidelberg, 2012, pp. 163173. isbn: 978-3-642-30023-3. doi: 10.1007/978-3-642-30023-3_15.
[4] A. Fischer. A special newton-type optimization method. Optimization 24.3-4 (1992), pp. 269284.
doi: 10.1080/02331939208843795. url: https://doi.org/10.1080/02331939208843795.
[5] Volker Fohrmeister, Alexander Bartels, and Jörn Mosler. Variational updates for thermomechan-
ically coupled gradient-enhanced elastoplasticity Implementation based on hyper-dual numbers.
Computer Methods in Applied Mechanics and Engineering 339 (2018), pp. 239261. issn: 0045-7825.
doi: 10.1016/j.cma.2018.04.047. url: http://www.sciencedirect.com/science/article/
pii/S0045782518302329.
325
[6] Christian Miehe and Jörg Schröder. A comparative study of stress update algorithms for rate-
independent and rate-dependent crystal plasticity. International Journal for Numerical Methods
in Engineering 50.2 (2001), pp. 273298. doi: 10.1002/1097- 0207(20010120)50:2<273::AID-
NME17>3.0.CO;2-Q.
[7] J. Mosler and O.T. Bruhns. On the implementation of rate-independent standard dissipative solids
at nite strain Variational constitutive updates. Computer Methods in Applied Mechanics and
Engineering 199.9-12 (Jan. 2010), pp. 417429. issn: 0045-7825. doi: 10.1016/j.cma.2009.07.006.
[8] Martin Schmidt-Baldassari. Numerical concepts for rate-independent single crystal plasticity. Com-
puter Methods in Applied Mechanics and Engineering 192.1112 (2003), pp. 12611280. issn: 0045-
7825. doi: 10.1016/S0045-7825(02)00563-7. url: http://www.sciencedirect.com/science/
article/pii/S0045782502005637.
[9] Masato Tanaka, Daniel Balzani, and Jörg Schröder. Implementation of incremental variational for-
mulations based on the numerical calculation of derivatives using hyper dual numbers. Computer
Methods in Applied Mechanics and Engineering 301 (2016), pp. 216241. issn: 0045-7825. doi:
10 . 1016 / j . cma . 2015 . 12 . 010. url: http : / / www . sciencedirect . com / science / article /
pii/S0045782515004156.
326

A thermomechanical finite element based framework for

phase transitions in laser based additive manufacturing
processes
1
Isabelle Guschke ∗
, Thorsten Bartel1 and Andreas Menzel1,2
1 Institute of Mechanics, TU Dortmund University, Dortmund, Germany
2 Division of Solid Mechanics, Lund University, Lund, Sweden
Abstract.
Additive manufacturing describes a unique processing technology, where a part is manufactured layer by layer
in contrast to traditional material removal or casting technologies. Within the laser based additive manu-
facturing processes, the metallic particles of a powder layer are molten by a laser beam. The industry has
a high interest in additive manufacturing processes, as custom-made light-weight parts can be produced for
a reasonable price. The influence of process parameters and predictability of the outcome are in so far still
challenging, as complex thermal and mechanical phenomena arise during the production cycle, whereby the
effects especially of high temperature gradients and phase transformations are difficult to predict. Thus, simu-
lations incorporating advanced material models are necessary to gain a better understanding of the process. In
this contribution, a physically well-motivated material model based on the minimisation of the free energy is
established for improved predictions of the effective material properties of the additively manufactured work-
piece. A thermomechanically fully coupled finite element based framework is employed to solve the balance of
linear momentum and the energy equation. With this at hand, it is possible to develop a more sophisticated
material model to gain a better understanding of the laser based additive manufacturing processes.
1 Thermomechanical framework
For the simulation of the additive manufacturing process, a fully coupled thermomechanical model is
used to capture the temperature evolution and the process-induced eigenstresses. The finite element
simulation is based on the balance of linear momentum and on the energy equation
∇·σ+b = 0 and (1)

− ∇ · q + r + Dmech + θ ∂θ [ σ : ε̇ − Dmech ] − e
c θ̇ = 0, (2)
respectively. In these equations, σ represents the stress tensor, b is the body force, q describes the heat
flux vector, r is the externally supplied heat, Dmech symbolises the mechanical dissipation, θ denotes the
temperature and e 2
c := − θ ∂θθ ψ defines the effective specific heat capacity based on the specific energy ψ.
For the time being, the linearised strain measure ε is considered appropriate.
2 Constitutive framework
The following model is adopted from phase transformations in shape memory alloys, see e.g. [1]. In
contrast to common models in additive manufacturing, i.e. [5], the underlying material model is based on
the definition of energy densities for each state of the material, i.e. the powder, molten and re-solidified
material. The goal is to develop a more advanced model which is based on constitutive relations for
each of the considered phases of the material rather than on the use of temperature dependent material
parameters.
∗ Corresponding author: isabelle.guschke@tu-dortmund.de ORCID iD: https://orcid.org/0000-0002-9841-6623
327
The composition of the material in terms of precise values for the phase fractions is obtained via energy
minimisation of the overall averaged mass specific energy density, i.e.
Z X
nph nph
X nph
X nph
1 ρi ξi V V X
ψ0 = ψi0 dmi = ζi ψi0 = ψi0 = ξi ψi , (3)
m i=1 i=1 i=1
m m i=1
where the mass specific energy of each phase is introduced as ψi0 and the volume specific energy as ψi
for all phases nph . The specific energies consist of a mechanical and a caloric part. The respective mass
fractions are denoted as ζi and the volume fractions as ξi . Furthermore, m defines the mass and V denotes
the initial volume.
It is assumed that the strain states in each phase εi may differ from each other, additively decompose
into elastic and inelastic parts and will minimise the energy, such that
Pnph
εi = arg min ψ 0 subject to r ε = i=1 ζi ε i − ε = 0 . (4)
Accordingly, the mass fractions ζi are determined via the minimisation while taking into account the
physical restrictions
Pnph
ζi = arg min ψ 0 subject to 0 ≤ ζi ≤ 1 and i=1 ζi = 1 . (5)
An additional constraint ζ̇pow ≤ 0 needs to be taken into account for the case of powder bed additive
manufacturing, as the rate of the powder fraction ζ̇pow cannot increase. Taking (4) and (5) into account,
the resulting effective energy obtained by the minimisation
ψ rel = min {ψ 0 } (6)

εi ,ζi
yields the relaxed energy density of the problem which is identical to the convex hull due to (4). The
effective stress σ can be calculated for the coupled thermomechanical problem as follows
∂ψ
σ= . (7)
∂ε
3 Representative examples
The energy densities of the phases determine the respective constitutive behaviour, whereas contributions
are chosen as
1 θ − θref
ψpow := εpow : Epow : εpow − cpow θ ln θ + Lpow
2
θref
ψmel := 21 εmel − εirr
mel : Emel : ε mel − ε irr
mel − cmel θ ln θ (8)
θ − θref
ψsol := 21 εsol − εirr irr
sol : Esol : εsol − εsol − csol θ ln θ + Lsol
θref
for the powder, molten pool and re-solidified phase, respectively. In these equations, Ei defines the
respective fourth-order isotropic elasticity tensor, εirri is the irreversible strain of the respective phase,
ci denotes the specific heat capacity and L represents the latent heat contribution with reference temper-
ature θref . Dissipative effects due to the irreversible behaviour of the molten and solid phase are captured
within Dmech .
For the molten phase and the solid phase, transformation strains εtrans
i are taken into account to capture
the volume changes of the material during phase transitions due to the different mass densities. In
addition, a viscoelastic strain εvis
mel is introduced to include the fluid-like behaviour of the molten pool
in the material model. In contrast, plastic deformation εplsol is used to capture the irreversible behaviour
of the solid phase. Furthermore, thermal strains εth
sol using a linear heat expansion model are considered
in the re-solidified phase. At a material point level, a representative boundary value problem based on
titanium aluminium alloy Ti6 Al4 V is implemented, see Figure 1 hereto. The following results are gained
while using slightly different densities.
As a first approach, equal densities in all phases are assumed so that conservation of mass is fulfilled a
priori, even if volume fractions are used as the conserved quantity, cf. [3]. For equal densities, the result
328
phase fractions ξmel • ζmel

1
strains εxx • εyy • εzz

ζmel
0 0.8 ξmel
0.6
-0.02
0.4
0.2
-0.04
0
0 1 2 3 4 5 0 1 2 3 4 5
time t [s] time t [s]
(a) Evolution of normal strain components. (b) Evolution of volume and mass fraction of
molten phase.
Figure 1. The temperature is prescribed at material point level. First, the temperature linearly increases from
θ = 1200 K to θ = 2200 K and then decreases back to θ = 1200 K again, while zero Neumann
boundary conditions are assumed. The resulting strain evolution can be seen in Figure (a), where the
severe changes occur due to the transformation strains. The linear evolution of strains after the full
transformation into the re-solidified phase is due to the linear heat expansion model. In Figure (b), the
evolution of the mass fractions, respectively volume fractions of the molten phase is shown, particularly
showing the difference between these two quantities.
of the convex hull defined by ψ rel coincides with the ansatz by the Reuss bound, where constant stress
states are assumed in each phase. For this case, the overall mass and volume specific energy density
coincide and the volume fractions match the mass fractions. Thus, no transformation strains εtrans i are
necessary. However, already for slight differences in the densities as they are applicable for the additive
manufacturing process, the sum of the volume fractions no longer equals one. These differences shall be
discussed in the following.
A small scale boundary value problem for a titanium aluminium alloy Ti6 Al4 V is implemented into
the commercial finite element software Abaqus to demonstrate the capacity of the constitutive model.
Here, the fully coupled thermal stress analysis is used which is based on the weak forms of (1) and
(2). The previously discussed constitutive material model is then implemented as a user material in the
subroutine UMAT. The related inequality constraints are solved by a regularised Fischer-Burmeister NCP
approach, see i.e. [2]. With this at hand, the temperature-induced change of the material’s composition
is obtained. The resulting mass and volume fractions are saved as internal state variables. In order to
finalise the overall material model, further homogenisation assumptions are applied to define the averaged
heat conductivity. A classic isotropic Fourier ansatz is adopted for the heat conduction model, and the
effective specific heat capacity is implemented in the subroutine UMATHT. For further information on the
implementation of a user material into Abaqus, see i.e. [6].
Additional challenges are related to the finite element based modelling of the manufacturing process
itself, e.g. in terms of a suitable modelling of the layer build-up and the adequate modelling of the laser
beam impact, as well as to choosing correct thermal and mechanical boundary conditions. The model
of the laser beam itself is implemented in the subroutine DFLUX, where the temperature field arises due
to a moving volumetric heat source, as introduced in [4] and applied to the simulation of selective laser
melting processes in, e.g. [3]. The resulting temperature and mass fraction evolution are shown for the
case of equal densities in Figure 2.
4 Conclusion
A thermomechancially fully coupled model is developed in this work for the simulation of laser based
additive manufacturing processes. A physically well-motivated material model based on the minimisation
of the free energy is established, where the different states of the material, namely powder, molten, and re-
solidified, are captured as single phases weighted by respective mass or volume fractions. Some numerical
examples for a small scale problem are used to illustrate capabilities of the overall framework.
329
Viewport: 1 ODB: /home/iguschke/Documents/.../Results/wo_Visco_RPL.odb
θ [K]
NT11
3005.74
1923.00
1768.00
1613.00
1458.00
1303.00
1148.00
993.00
838.00
683.00
528.00
373.00
(a) Temperature distribution θ.

Viewport: 1 ODB: /home/iguschke/Documents/.../Results/wo_Visco_RPL.odb
SDV2
ζmel
1.37
1.00
0.90
0.80
0.70
0.60
0.50
0.40
0.30
0.20
0.10
0.00
-0.36
(b) Evolution of mass fraction ζmel for the molten phase for equal densities.
Figure 2. A representative example is used to demonstrate the additive manufacturing process using equal densities
for all phases. These plots show in Figure (a) the temperature distribution and in Figure (b) the mass
fraction of the molten pool ζmel at an identical time step induced by the moving volumetric heat source
modelling the laser beam impact.
References
[1] T. Bartel and K. Hackl. A micromechanical model for martensitic phase-transformations in shape-
memory alloys based on energy-relaxation. Zeitschrift für Angewandte Mathematik und Mechanik
89.10 (2009), pp. 792–809. doi: 10.1002/zamm.200900244.
[2] T. Bartel, R. Schulte, A. Menzel, B. Kiefer, and B. Svendsen. Investigations on enhanced Fischer-
Burmeister NCP functions: application to a rate-dependent model for ferroelectrics. Archive of Ap-
plied Mechanics 89.6 (2019), pp. 995–1010. doi: 10.1007/s00419-018-1466-7.
[3] T. Bartel, I. Guschke, and A. Menzel. Towards the simulation of Selective Laser Melting processes via
phase transformation models. Computers & Mathematics with Application (available online 2018).
doi: 10.1016/j.camwa.2018.08.032.
[4] J. Goldak, A. Chakravarti, and M. Bibby. A New Finite Element Model for Welding Heat Sources.
Metallurgical Transactions B 15 (1984), pp. 299–305. doi: 10.1007/BF02667333.
[5] D. Riedlbauer, P. Steinmann, and J. Mergheim. Thermomechanical finite element simulations of
selective electron beam melting process: Performance considerations. Computational Mechanics 54
(2014), pp. 109–122. doi: 10.1007/s00466-014-1026-0.
[6] Dassault Systèmes. Abaqus Analysis User’s Manual v2016. 2015.
330

A Sliding Interface Approach with Application to Twin-

Screw Extruders
1
Jan Helmig ∗
and Stefanie Elgeti1
Abstract.
Co-rotating twin-screw extruders are important processing devices within the plastics-producing industry.
The utilized screw elements are conveying and kneading elements. For conveying elements the geometry
can be considered smooth in longitudinal direction. In contrast, kneading elements feature discontinuities in
longitudinal direction. Within this work we present a finite-element framework to solve the flow of plastic
melt inside extruders. For such simulations the methods presented in [2, 3] allow to use boundary-conforming
meshes for conveying elements. However, kneading elements with their discontinuities in longitudinal direction
remain out of scope. Thus, we introduce a sliding interface approach based on Nitsche’s method [5] combined
with a multimesh technique [4] that couples the solution fields of the individual meshes at the interface.
In contrast to standard fictitious domain methods, cut cells only have to be computed in 2D which avoids
difficulties with degenerated polygons. We show the potential of the method for a testcase that computes the
flow of plastic melt in a realistic extruder section.
Introduction
Co-rotating twin-screw extruders are important processing devices within the plastic-producing industry:
Reasons are their versatile modes of usage in processing, mixing, and reaction. The utilized screw elements
can be roughly classified into either conveying and kneading elements, Fig. 1(a). The design of conveying
elements is based on a sweep of a 2D cross section, see Fig. 1(b). The geometry can be considered smooth
in longitudinal direction. Instead, kneading elements feature geometrical discontinuities in longitudinal
direction, see Fig. 1(a).
(a) 3D extruder section including kneading (b) 2D extruder cross section.

and conveying elements.
Figure 1. Sketch of a twin-screw extruder.
Flow simulations of plastic melts inside twin-screw extruders pose exeptional difficulties. This is due to
the constantly rotating screws in combination with small gap sizes. Furthermore, modeling the plastic
melt is a multi-physics problem: The melt is heated up due to shearing, which then in turn influences
the material properties of the melt. Thus, the flow equations have to be coupled to the heat equation. In
∗ Corresponding author: helmig@cats.rwth-aachen.de
331
terms of numerical approach, we aim at using a boundary-conforming finite-element method within this
work.
Boundary-conforming meshes have two major advantages: (1) It is straightforward to strongly impose
the kinematics of the screw, and (2) it is easy to construct high-quality boundary-layer meshes. The
challenge arises when those meshes need to be deformed; especially without re-meshing. A suitable
deformation method – the Snapping Reference Mesh Update Method (SRMUM) – has been presented
in [2]. It already enables the simulation of unsteady flow inside convex-shaped conveying elements. A
method also amenable to concave shapes has been presented in [3]. It extends the idea of SRMUM with a
spline-parameterization of the geometry. Still, kneading elements with their discontinuity in longitudinal
direction remain out of scope. The methods described are able to generate and update high-quality
boundary-conforming meshes for each individual continuous part of a kneading block. However, these
individual meshes do not match at the interface.
The major part of this work aims to couple the individual meshes at their interfaces. We present a sliding
interface approach that is inspired by [1, 5]. It allows to couple two meshes at an interface, composed of
parts of the mesh boundary surfaces. Breaking down the coupling to a surface is advantageous since this
simplifies the computation of the combined interface triangulation, cut cells, and avoids difficulties with
degenerated polygons.
1.1 Governing equations and general solution method
The behavior of plastic melt inside an extruder is modeled as the flow of a viscous incompressible
temperature-dependent fluid. It is governed by the incompressible Navier-Stokes equations. We consider
Newtonian as well as Generalized-Newtonian models for the constitutive equation. For Generalized-
Newtonian models the viscosity of the plastic melt is not constant anymore. It depends on shear rate
and temperature and thus enables to account for shear-thinning effects. We employ a semi-discrete finite
element method in order to solve the governing equations. We use first order interpolation for all degrees
of freedom. Thus, we have to stabilize our equations using a Galerkin-Least-Squares approach. Further-
more, the rotation of the screws results in a constant movement of the computational domain and thus,
the mesh. In order to account for that movement we employ the Arbitrary-Lagrangian-Eulerian (ALE)
approach. Additionally, we decouple the temperature equation from equations governing conservation
of mass and momentum. Still, the flow and temperature field are coupled strongly using a fixed-point
iteration.
1.2 Sliding interface approach
Within this section, we will introduce the most important ideas of the sliding interface approach that
can be used to couple the solution fields of two boundary-conforming meshes at a common interface Γf f .
We refer to this method as sliding interface since the meshes are allowed to move/rotate and thus slide
over each other. A sketch of Γf f for an exemplary composition of two kneading elements is given in Fig.
2. In Fig. 2(a) the cross sections of the two individual boundary-conforming meshes are given as well as
the interface itself in Fig. 2(b). The interface Γf f can be seen as Γf f = Γred ∩ Γblue . In the general case
we consider Γf f = Γ1 ∩ Γ2 .
The motivation of the method was given in [1], where rotating domains are coupled with steady ones based
on a common spline interface. In the following we will neglect the temperature for simplicity. However,
the concept for the sliding interface approach for the temperature field works in the same way as for the
flow field. We refer to the combined discrete stabilized variational formulation for both individual meshes
without the coupling terms as B({v h , qh }, {uh , ph }), where uh , ph are the discrete velocity and pressure
fields and v h , qh the corresponding test functions. In order to account for the coupling we add Nitsche
coupling terms at the common interface and obtain the combined discrete variational formulation:
332
(a) 3D view of the interface. (b) 2D cross section. Γf f is composed of the two interfaces
Γred and Γblue .
Figure 2. Example kneading element.
B N itsche ({v h , qh }, {uh , ph }) = B({v h , qh }, {uh , ph })

+ hJv h K · n, {ph }iΓf f − hJv h K, {2ηε(uh ) · n}iΓf f − h{qh }, Juh K · niΓf f − h{2ηε(v h ) · n}, Juh KiΓf f
1
+ hγη/hn Jv h K, Juh KiΓf f − hρ(uh · n){v h }m , Juh KiΓf f + hρ |uh · n| Jv h K, Juh KiΓf f = 0.
2
(1)
The terms in the second and third line of Eq.(1) weakly enforce the continuity of velocities and tractions
on the interface [5]. The third last term penalizes the discrete version. In order to evaluate the integrals
introduced, it is necessary to compute all cuts of the two discretizations at the interface. We subdivide
the resulting cut elements into triangles to define quadrature rules.
So far, the presented approach neglected possible movements/obstacles at the interface. Looking at Fig.
2(b) we want ensure that e.g. flow in normal direction at Γred is only allowed on Γf f . Furthermore, for
Γred Γf f we attempt to apply a no-slip condition on the neighbouring screw. We enforce this condition
g, weakly. This is done by enhancing B N itsche ({v h , qh }, {uh , ph }):
B SlidInt ({v h , qh }, {uh , ph }) = B N itsche ({v h , qh }, {uh , ph })

X 2 h
+ (v h · n, ph )Γi Γf f − (v h , 2ηε(uh ) · n)Γi Γf f
i=1
i
− (qh , (uh − g) · n)Γi Γf f − (2ηε(v h ) · n, (uh − g))Γi Γf f + (γη/hn v h , (uh − g)Γi Γf f = 0. (2)
One open question is how to determine Γi Γf f . The construction requires that we detect all elements
that are not cut or only partly cut. This might result into cut polygons that are not necessarily convex.
In order to avoid this extra computational effort, we use the multimesh concept introduced in [4]. The
idea is to apply the extra terms added in Eq.(2) on Γi and subtract them on Γf f by using negative
quadrature weights.
In a first step, the presented sliding interface approach has been used to compute the flow of Newtonian
fluids inside a simplified kneading block with only two kneading elements for a single-screw extruder. The
computational domain is given in Fig. 3(a). Fig. 3(b) shows the velocity distribution in axial direction
for the two non-matching interfaces. The validity of the weak imposition of boundary conditions is clearly
visible since the axial velocity is zero in that part of each interface that is attached to the opposite screw.
Furthermore, the velocities match at both interfaces. This clearly indicates that the method correctly
couples the flow throughout the non-matching interface.
333
(a) 3D flow results. (b) z-velocity distribution at the non-matching interface.
Figure 3. Single-screw extruder with one kneading section.
Figure 4. z-velocity distribution inside a twin-screw extruder section including conveying and kneading elements.
In the following, we use the presented method to compute the flow of a shear-thinning fluid inside a
realistic twin-screw extruder section. It consists of forward and backward conveying as well as kneading
elements. Fig. 4 shows the axial velocity on a slice in the xz-plane. The transition between non-matching
meshes is very smooth.
3 Conclusion
We have presented a sliding interface approach that enables to couple two boundary-conforming meshes
at a common moving interface. It is based on Nitsche’s method combined with a multimesh approach.
This simplifies the computation of cut cells at the interface. We show flow results of plastic melts inside
single- and twin-screw extruders that demonstrate the validity of the proposed method.
References
[1] Yuri Bazilevs and Tom Hughes. NURBS-based isogeometric analysis for the computation of flows
about rotating components. Computational Mechanics 43.1 (2008), pp. 143–150.
[2] Jan Helmig, Marek Behr, and Stefanie Elgeti. Boundary-Conforming Finite Element Methods for
Twin-Screw Extruders: Unsteady - Temperature-Dependent - Non-Newtonian Simulations. Comput-
ers & Fluids (2019).
[3] Jochen Hinz, Jan Helmig, Matthias Möller, and Stefanie Elgeti. Boundary-Conforming Finite El-
ement Methods for Twin-Screw Extruders using Spline-Based Parameterization Techniques. arXiv
preprint arXiv:1905.11829 (2019).
[4] August Johansson, Benjamin Kehlet, Mats G Larson, and Anders Logg. Multimesh finite element
methods: Solving PDEs on multiple intersecting meshes. Computer Methods in Applied Mechanics
and Engineering 343 (2019), pp. 672–689.
[5] Andre Massing, Benedikt Schott, and Wolfgang A Wall. A stabilized Nitsche cut finite element
method for the Oseen problem. Computer Methods in Applied Mechanics and Engineering 328 (2018),
pp. 262–300.
334

Partitioned coupling of fluid-structure interaction for

the simulation of floating wind turbines
Jorrid Lund∗1 , Bjarne Wiegard1 , Lars Radtke1 , Marcel König1 and Alexander Düster1
1 Numerical
Structural Analysis with Application in Ship Technology (M-10), Hamburg University of Technology, Am Schwarzen-
berg-Campus 4c, 21073 Hamburg, Germany
Abstract. Fluid-structure interaction problems appear in many different fields of application. An interesting
and highly topical field is offshore renewable energy. An increasing energy demand, limited fossil energy
sources and the climate change in combination with limited suitable and available space ashore lead to an
increased relevance of offshore renewable energy. A promising offshore renewable energy concept is the use of
floating wind turbines.
For the simulation of floating wind turbines, the necessary and computationally expensive fluid-structure
interaction coupling has to be combined with large domains, highly nonlinear and dynamic structural behavior
and complex flow situations. This requires new and innovative techniques in the solution process.
In this work, the partitioned solution utilizing existing field solvers is suitable for this type of problem due to
the availability of well-developed fluid and structural solvers.
To manage the partitioned coupling, the in-house C++ library comana is employed. This coupling framework
is employed to control the iterative solution process and to perform the necessary exchange of surface quantities
on the interface between the fluid and structural field of the fluid-structure interaction problem. Through this,
comana enables the computation of complex strongly coupled fluid-structure interaction problems within a
reasonable amount of time. Despite the necessary changes for the partitioned coupling, the modifications to
the existing specialized high-fidelity fluid and structural solvers are minimized to maintain their sophisticated
solution techniques.
The partitioned coupling procedure is described and simulation results are presented.
Different measures to accelerate, stabilize and improve the solution of this fluid-structure interaction problem
are demonstrated. Among these measures are the prediction and the convergence acceleration in the implicit
coupling of the fluid and structural field solvers.
In the scope of a future research project, the presented method will be applied to develop the simulation of a
point wave energy absorber used for the conversion of wave energy into electric energy.
Introduction
The partitioned coupling strategy is a powerful solution approach for many fluid-structure interaction
(FSI)-problems. Due to the partitioned approach, existing specialized field solvers can be used to solve
FSI-problems only with the exchange of surface quantities in a fast and efficient way. Among the many
applications of this approach are offshore renewable energies of which the most prominent technology are
offshore wind turbine plants. Wind blows stronger and more continuously offshore, which is an important
advantage of offshore wind energy turbine plants especially with respect to power grid stability in face
of the transition to renewable energies. Even though offshore wind energy has many advantages over
traditional onshore wind energy plants, it also poses serious challenges. For large water depths, offshore
wind energy turbine plants with a foundation in the seafloor are not cost-effective. For these water
depths, floating wind turbines are more suitable and economic. The construction and design of floating
wind turbines, however, poses serious challenges since the forces on the structure, which are affected by
the dynamic motion and structural deformations, are hard to determine with uncoupled structural or
fluid simulations. A coupled simulation nevertheless requires well developed solvers able to cover the
different physical effects that have to be taken into account for such a complex simulation.
∗ Corresponding author: Jorrid.Lund@tuhh.de ORCID iD: 0000-0003-4537-8720
335
1 Partitioned solution approach for FSI-problems
For the solution of FSI-problems, fluid and structural equations, which are dependent on each other, have
to be solved in a coupled approach. For this, it is possible to solve both subproblems by coupling them
in a partitioned (segregated) or in an monolithic (integrated) way [felippa].
To solve the fluid and structural equations in an integrated way, the monolithic approach requires the
development of solvers appropriate to the given FSI-problem. Integrated FSI solvers are more robust, but
require the labor-intensive development of a dedicated solver for the problem at hand [bazilev]. They
are not easily modifiable in the selection of the discrete equations that describe the structural or fluid
problem [bungartz]. Partitioned solution procedures on the other hand allow for an independent and
flexible modelling to select the most appropriate techniques for the solution of the structural or fluid side.
This enables the usage of highly specialized solvers for both subproblems in a black-box manner, where
only the surface quantities have to be exchanged [bungartz].
The in-house software library comana has been developed to control this iterative exchange of quantities
in multi-physic problems [Koenig].
To compute the displacements and velocities of a solid body under external loads, the finite-element-
method (FEM) is among the most popular discretization schemes. For the computation of the flow field
in a moving domain, finite-volume-method (FVM) and boundary-element-method (BEM) are widely used
methods to compute the pressure and velocity inside the fluid domain.
In order to couple the FEM with the BEM or FVM to account for FSI, the displacements d~ calculated
by the structural solver Ss are interpolated to the nodes on the boundary of the fluid solver such that
the fluid mesh can be deformed according to the movement of the structure.
The fluid solver Fs can then compute tractions t~i = Fs (d~i ) on the interface between the fluid and
structural side in the deformed domain. These tractions can be applied to the structure as external loads
to compute new displacements d~i+1 = Ss (t~i ).
This procedure can be seen as a fixed-point iteration since the quantity (in this case displacements) that
is computed is the same that enters the procedure:
d~i+1 = Ss (Fs (d~i )) (1)
The procedure described by equation (1) is repeated until the displacement increment between two
subsequential iterations reaches a predefined tolerance. The converged displacements are taken as the
solution for the current time step and the structural and fluid solvers are advanced in time such that the
procedure can be repeated for the next time step.
In order to speed up and stabilize the fixed-point iteration (1), different measures can be taken.
The prediction of the displacements at the beginning of a time step is used to provide a good initial guess
for the displacements entering the fixed-point iteration. A polynomial of order p depending on time t can
be constructed based on the previous m = p + 1 time steps by requiring that the polynomial is equal to
the displacements at the previous m time steps.
p
X
˜ j) =
d(t ai tij for j = −1, ..., −m (2)
i=0
The resulting system of equations (2) can be used to determine the parameters ai . With these parameters
ai the prediction of the displacements for the current time step d(t˜ 0 ) can be computed based on the
previous time steps. The dependence on the previous time steps is constant in the case of a constant
time step size.
While the prediction operates on the solution of previous time steps, the results within the implicit
iterations of one time step can also be improved based on the solution of previous iterations in the
same time step. To accelerate and stabilize fixed-point iterations, a variety of methods exist. The
simplest method is the constant relaxation which just multiplies a constant factor to the increment
between two iterations. For FSI-problems, advanced methods like Aitken or quasi-Newton methods are
336
more suitable. Among quasi-Newton methods, the quasi-Newton least square method has shown good
convergence behaviour for many FSI-problems [Radtke][Degroote].
2 Floating wind turbine
To demonstrate the benefits of the simulation of FSI using partitioned coupling, complex real world
problems like the simulation of a floating wind turbine is well suited. The floating wind turbine presented
here was developed within the joint research program ‘Hydrodynamic and Structural Optimization of a
Semi-submersible Offshore Wind Turbine’. The related FSI simulations were carried out in the subproject
‘Fluid-Structure Interaction and Optimization of a Floating Wind Turbine’.
Figure 1. Schematic overview Figure 2. Stresses inside the structure and elevation of the water surface
[phdkoenig]
In Figure 1 an overview of the floating wind turbine is given. This floating wind turbine consists of a
platform to lift and stabilize the whole structure and a tower on which the rotor is mounted. To prevent
drifting, the platform is anchored in the seabed such that it is self-aligning in the wind [Wiegard].
The deformation of the platform, the tower and the rotor is simulated with FEM using the commercial
software package ANSYS [Ansys]. The fluid simulation is performed using the boundary element method
implemented in the in-house software panMARE [Netzband]. comana [Koenig] is used to control the
iterative solution procedure and exchange tractions and displacements between ANSYS and panMARE.
In Figure 2 the result of a coupled FSI simulation is shown. Not only the forces on the tower and rotor
due to the wind but also the effects of the ocean current and waves is taken into account. Suitable
techniques like sub-structuring with large rotations [Wiegard] available in the structural solver can be
used without modifying the partitioned coupling procedure.
The partitioned simulation allows for the evaluation of deformations and stresses at all points in the
structure in every time step taking the important dynamic and nonlinear effects into account. Not only
the structural integrity of the platform but also the orientation of the platform in the wind and waves
can be evaluated.
The partitioned approach turned out to be suitable for FSI simulations. In a recent research project,
the motion and the structural behaviour of a floating wind turbine was successfully investigated. A
precise stress analysis is possible and effects of the structural deformation on the motion behaviour can
be evaluated [Wiegard].
Even though the simulation of the floating wind turbine showed good results, the boundary element
method can not fully account for the flow separation and turbulence in the wake of the rotors. To
337
increase the solution accuracy, the region where turbulence effects and flow separation occur can be
modelled by the FVM, which is more accurate but at the same time computationally more expensive.
The FVM can be applied to the wake region of the rotor while the outer region which exhibits non of
these effects could be modelled with BEM.
The techniques which have been applied to the floating wind turbine can also be transferred to other
floating offshore plants. An example for such an application is the point wave energy converter. The
approach is promising for every type of floating structures which is subjected to external forces.
Acknowledgments
The research regarding the floating wind turbine was funded by the Federal Ministry for Economic Affairs
and Energy of Germany (BMWi) in the joint project HyStOH (project number 03SX409C).
338

Particle Finite Element Simulation for Additive Manu-

facturing Processes of Fresh Concrete
Janis Reinold
∗ 1, Jithender Jaswant Timothy
1 and Günther Meschke
1
1 Institute for Structural Mechanics, Ruhr-University Bochum
Abstract. In this work a computational model based on the Particle Finite Element Method for the simulation
of extrusion-based additive manufacturing of fresh concrete is presented. The constitutive law is based on a
Bingham model, which is approximated by a classical elasto-viscoplastic Perzyna model in a rate form. A
mixed velocity-pressure formulation in an updated Lagrangian framework is chosen to solve the system of
equations. Model capabilities are demonstrated using two numerical test cases.
Introduction
Additive manufacturing techniques so far have found application predominantly in the manufacture of
high-end components in the bio-medical and aeronautical industry mainly due to the high costs of the
materials used and the specialized technology associated with the technique. Recently there is grow-
ing interest in the development of this technology for large scale fabrication of concrete structures [1].
The most common approach is based on layered extrusion of fresh mortar or concrete, that allows for
automated construction of concrete structures and components. Due to process or machine parameters
and the time dependent nature of (fresh) concrete, the nal properties of printed components are much
more dependent on the 'casting' process, than in conventional casting of concrete. Therefore, numerical
models can help to understand the complex interactions of various parameters in additive manufacturing
processes of fresh concrete. To this end, a numerical framework based on the Particle Finite Element
Method (PFEM) is developed for the simulation of extrusion-based additive manufacturing processes of
fresh concrete. Flow of homogenized fresh concrete is generally based on the Bingham model, which is a
non-Newtonian yield stress uid. Hence, a certain yield stress must be exceeded to initiate viscous ow.
In many engineering problems the material response before yielding is neglected and assumed to be rigid,
which is retained by regularized models and a large viscosity before yielding. These models have the
advantage that no internal variable must be stored. However, a drawback of these methods is, that the
stresses of the material at 'rest' are deduced from the material that never really comes to rest. There-
fore, the Bingham model is modeled with a classical elasto-viscoplastic model, so that the viscoplastic
contribution is approximated with an overstress function. In contrast to regularized rigid-viscoplastic
Bingham models, stresses of material at rest can be assessed more accurately, which is relevant for the
characterisation of initial stress states of printed components in additive manufacturing processes of fresh
concrete.
1.1 Constitutive law
In order to describe the ow of homogenized fresh concrete, the constitutive law is based on a Bingham
model, which is approximated by a classical elasto-viscoplastic Perzyna formulation [4]. The elastic parts
of the strains are assumed small in comparison to the plastic deformations. Under these assumptions, the
material model is given in a hypoelastic rate form and the deformation rate tensor d is additively split
∗ Corresponding author: janis.reinold@rub.de
339
into elastic and plastic parts as d = de + dp with de and dp denoting the elastic and plastic deformation
rates, respectively. The objective Jaumann stress rate of the Kirchho stresses is given as
τ ∇J = C : (d − de ) (1)
with the the elasticity tensor C = 2µ(I − 31 I ⊗ I) + κI ⊗ I where µ denotes the shear modulus, κ the
bulk modulus, Iij = δij the second order identity tensor and Iijkl = (δik δjl + δil δkj )/2 the fourth order
symmetric identitiy tensor. In comparison to Bingham models based on regularized formulations, a xed
yield function need to be satised as
Φ(τ ) = q(τ ) − σ0 (2)
q
with q(τ ) = 2 dev[τ ] : dev[τ ] denoting the equivalent stress and σ0 the yield stress. As a standard
3
procedure from computational plasticity, the plastic ow rule [5] is dened by
∂Φ
dp = γ̇ (3)
∂τ
in which γ̇ is the plastic multiplier, that must obey the standard Kuhn-Tucker loading/ unloading condi-
tions for classical rate-independent plasticity Φ(τ ) ≤ 0, γ̇ ≥ 0 and Φ(τ )γ̇ = 0 along with the consistency
condition γ̇ Φ̇(τ ) = 0. The Bingham model is obtained by a Perzyna formulation [4] of the plastic
multiplier with an overstress function equal to the yield function as
< q(τ ) − σ0 >
γ̇ = (4)
η
where η is a viscosity parameter, which is related to the uid viscosity µ as η = 3µ and "< . >" are the
McCauley brackets given as (
f (τ ) if f (τ ) ≥ 0,
< f (τ ) >= (5)
0 if f (τ ) < 0.
1.2 Governing equations
The local form of the balance of momentum for homogenized fresh concrete, covering the domain Ω in
the time interval (0, T ), becomes
Dv
ρ =∇·σ+b in Ω × (0, T ) (6)
Dt
where ρ denotes the density, v the velocity, σ the Cauchy stress tensor b the body forces. The boundary
conditions on the Neumann boundary ∂ΩN are σ(x, t)n = tp (x, t) with the normal vector n and the
prescribed tractions tp (x, t). The boundary conditions on the Dirichlet boundary ∂ΩD are v(x, t) =
v p (x, t) with the prescribed velocities v p (x, t). Due to the form of the underlying plasticity model,
situations may appear, where the material response is nearly incompressible. Hence, locking phenomena
might occur with a pure velocity formulation. Therefore, the mean stress is separated from the stress
eld and discretized over an additional equilibrium equation, the balance of mass. The balance of mass
is dened by
1 Dp
∇·v = in Ω × (0, T ) (7)
κ Dt
with the mean stress or pressure p and the bulk modulus κ. Consequently, the bulk modulus κ is also
separated from the elasticity tensor and objective stress integration is just performed on the deviatoric
parts of the stress tensor.
2 Numerical framework and the particle nite element method
The governing equations (6) and (7) are solved using the Particle Finite Element Method (PFEM) [2].
In PFEM the underlying weak form is discretized by means of standard nite elements in an updated
Lagrangian formulation. The computational domain is frequently re-meshed to avoid element distortion.
Hence, fast remeshing is essential and is achieved by using triangular and tetrahedral elements in two or
three dimensions, respectively. The remeshing algorithm consists of two steps: First a standard Delaunay
340
triangulation of a point cloud is formed and second, in order to describe the free surface of the domain,
elements with a circumcircle or circumsphere larger than a threshold value α are removed from the mesh.
This procedure is called the α-shape method.
In the context of this work, the mixed velocity-pressure formulation is discretized by linear nite elements
and the formulation is stabilized by a direct pressure stabilization.
3 Numerical application: Dam break test
The ow behavior of the proposed numerical model is compared with the ow of a regularized Bingham
model [3] using a typical dam break test. See Figure 1a) for the geometry of the problem. The material
and simulation parameters are the plastic viscosity µ = 5 Pa · s, the dynamic yield stress µ = 300 Pa, the
3
density ρ = 2000 kg/m and the time step size ∆t = 2·10−5 s. The study is performed with three dierent
element sizes h = 0.01, 0.005, 0.0025. Finally, the free surface of the material for the nest element size
is given for dierent time instances in Figure 1b). As can be observed, the time dependent ow behavior
is almost identical to the one obtained from a regularized Bingham model. Similar observations can be
made from the evolution of the current material front position over time, see Figure 1c). Furthermore,
as can be observed from Figure 1c), the ow of the material stops at around 0.6 s.
0.1 m
Fluid front position [m]

Perzyna model
Y-coordinate [m]
0.15 0.5
regularized Bingham model
0.4
full slip 0.1 t=0.2 s t=0.35 s
Perzyna model h=0.01
final 0.3
0.3 m
Perzyna model h=0.005

0.05 Perzyna model h=0.0025
0.2
regularized Bingham model
no slip 0 0.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
a) b) X-coordinate [m] c) Time [s]
Figure 1. a) Geometry of the dam break test b) Shape of the material for dierent time instances of the dam
break test c) Fluid front position over time diagram for the dam break test
4 Numerical application: 3D-concrete-printing
5,8 cm t=2.5 s
Extrusion
Inlet face nozzle
t=7.5 s
vinlet vprint
t=12.5 s
~2,7 cm
t=15 s
50 cm
Figure 2. a) Geometry of the extrusion-based additive manufacturing of concrete scenario b) Printed shapes of the
extrusion-based additive manufacturing test case scenario for dierent time instances with τ0s = 1 kPa
In this scenario of 3D-concrete-printing, three layers of 50 cm length and 2.7 cm height are numerically
printed, see Figure 2a). The simulation is conducted in 3D by using symmetry along the printing direction.
The presented results focus on the shape of the concrete after leaving the extrusion nozzle and on pressure
acting on lower layers, which is generated by the extrusion of the material. Due to a short process time
of a few seconds, the inuence of time dependent material behavior (structural build-up) is neglected in
this analysis. The ow rate of the extruded concrete at the nozzle outlet is assumed to be constant over
the cross section. Hence, the interaction between processes in the machine and the resulting ow rates is
not considered. Therefore, the inlet ow from the extrusion nozzle is approximated with a velocity inlet,
see Figure 2a). The material and simulation parameters are viscosity µ = 15 Pa·s, dynamic yield stress
3
τ0 = 0.75 kPa, density ρ = 2000 kg/m , Young's modulus E = 0.1 MPa, Poisson's ratio ν = 0.3, printing
velocity vprint = 0.1 m/s, ow rate vinlet = 0.09744 m/s, nozzle-separation distance ∆znozzle = 2.7 cm,
∆t = 0.5 · 10−4 s and tmax = 15 s. For reecting dierent states of structuration, the static yield stress
341
of already printed layers is set to a value of τ0s = 1, 1.5 and 2 kPa. This yield stress is responsible to
prevent the material from collapse in the printing process.
The numerically printed shapes of the test case scenario of extrusion-based additive manufacturing of
fresh concrete with a static yield stress τ0s = 1 kPa for dierent time instances are depicted in Figure 2b).
In Figure 3a) the outline of the cross section of each additively manufactured concrete layer are shown.
The results are given for the time instances when the rst layer (t = 5 s), the second (t = 10 s) and third
layer (t = 15 s) is printed. As can be observed, there is a signicant squeezing of lower layers in the
manufacturing process, when material is extruded on top of already printed layers. Fresh concrete slightly
ows in the direction perpendicular to the printing direction. These observations are underlined by the
results depicted in Figure 3b), where the yielded (red) and unyielded (blue) regions of the manufacturing
process for the time instance t = 12.5 s are given. Figure 3b) reveals that the equivalent stress exceeds
the yield stress in a certain region below the extrusion nozzle of lower layers. Hence, the material is not
strong enough to bear the pressure generated from the inlet ow. By changing the static yield stress to
values of τ0s = 1.5 and 2 kPa for reecting dierent states of structural build-up and a certain waiting
time between each layer, lower layers become strong enough to carry the load from the inlet ow, see
Figure 3b).
0.1
t=5 s
0.08 t=10 s
t=15 s �0s =1 kPa
Height [m]
0.06
0.04
�0s =1.5 kPa
0.02
0
-0.05 0 0.05 �0s =2 kPa
a) Width [m] b)
Figure 3. a) Outline of the cross section of the additively manufactured concrete layers for dierent time instances
t = 5 s (rst layer printed), t = 10 s (second layer printed), t = 15 s (third layer printed) with τ0s = 1 kPa
b) Yielded (red) and unyielded (blue) regions for dierent static yield stresses for t=12.5 s (view on the
symmetry line)
5 Conclusions
A computational model based on the Particle Finite Element Method for the simulation of extrusion
processes of 3D-concrete-printing techniques is presented. The constitutive law is based on a Bingham
model, which is approximated with a classical elasto-viscoplastic formulation. Results show that the ow
behavior of this model is almost identical to the ow behavior of models with a regularized Bingham
model. Furthermore, robustness of the proposed model is shown in a numerical test case scenario of
extrusion processes in additive manufacturing of fresh concrete. Simulations of a 3D-printing process
show that in addition to the self weight of upper layers, the extrusion process itself generates a pressure
on lower layers, that temporarily also need to be carried by these layers. If the static yield stress of these
layers is not large enough, the yield stress is exceeded and lower layers are squeezed, which results in
unintended displacements that might adversely inuence the integrity of the printed layers during the
additive manufacturing process.
References
[1] S. Lim, R.A. Buswell, T.T. Le, S.A. Austin, A.G.F. Gibb, and T. Thorpe. Developments in construction-
scale additive manufacturing processes. Automation in Construction 21 (2012), pp. 262 268.
[2] Eugenio Oñate, Sergio R Idelsohn, Facundo Del Pin, and Romain Aubry. The particle nite element
method: an overview. International Journal of Computational Methods 1.02 (2004), pp. 267307.
[3] T.C. Papanastasiou. Flows of Materials with Yield.Journal of Rheology 31.5 (1987), pp. 385404.
[4] P. Perzyna. Fundamental problems in viscoplasticity. In: Advances in Applied Mechanics. Ed. by
G. Kuerti. 9. Academic Press, New York, 1966, pp. 244368.
[5] J.C. Simo and T.J.R. Hughes. Computational Inelasticity. Berlin: Springer, 1998.
342

Optimisation of Diffusion Driven Degradation Processes

1
Navina Waschinsky ∗
, Franz-Joseph Barthold1 and Andreas Menzel2,3
1 Structural Mechanics, TU Dortmund, Dortmund, Germany
2 Instituteof Mechanics, TU Dortmund, Dortmund, Germany
3 Division of Solid Mechanics, Lund University, Sweden
Abstract. It is known that nature always strives for optimal processes in order to prevent early damage or
failure, for example bones which regulate the tissue in high stress areas or the optimal shape of trees. Many
materials in nature are highly complex as they are designed for long duration. Taking the nature itself as
a role model means taking its complexity into account, especially when considering materials which suffer
from chemical attacks and the degradation of materials. From this point of view, we create an algorithm
which considers degradation processes triggered by chemical substances which depend on diffusion conditions
in a finite element (FE) framework. In addition, we perform design optimisation with stress restrictions by
taking evolutionary processes as a role model. In this paper a biological inspired algorithm which consid-
ers degradation processes triggered by chemical substances is presented. We create a coupled diffusion and
growth/degradation approach with a permeable structure allowing a gradient-based flow of concentrations.
The continuum describes a one-phase approach with continuum modeling of a growth/degradation process.
With a thermodynamically consistent evaluation we formulate the govering equations, including stress, particle
diffusion and chemical reactions.
Introduction
This paper focuses on the general outline of the algorithm combining a specific structural coupled problem
to an optimisation framework. In detail, we deal with the coupling of diffusion and growth/degradation
processes embedded in a design optimisation algorithm.
In the following we present the evaluation of the structural analysis leading to a fully coupled diffusion
and growth/degradation material. We outline the kinematic approach of isotropic growth combined with
the constitutive evaluation of the chemical particles which are the trigger for mass growth or degradation.
Furthermore, we present the thermodynamically consistent evaluation of the govering equations, including
mechanical stress, particle diffusion and chemical reactions. The structural analysis is implemented in a
Fortran code, which is embedded in a Matlab framework. The advantages of the power of Matlab allow
the application of a Matlab optimisation toolbox within a finite element method (FEM) formulation. We
outline the general optimisation algorithm referring to a simple example.
1 Structural Analysis
1.1 Kinematics
The kinematic concept enables the description of all tangent operations in the continuum mechanical
concept, for further information see [1, 2, 3]. Figure 1 gives a short overview of the assumptions. We
introduce the coordinates in the reference configuration Xi , in the actual configuration xi , in the growth
space Xb i and in the parameter space Θi , namely the convective coordinates. Furthermore, we apply
mappings between the configurations for geometry κ, for motion µ and for growth γ. Partial derivatives
of the mappings with respect to the respective local coordinates result in the convective tangent vectors Gi ,
gi and Gb i . The gradients of the mappings between the parameter space and the introduced configurations
lead to the local, fundamental gradients K, M and T.
∗ Corresponding author: navina.waschinsky@tu-dortmund.de ORCID iD: 0000-0002-4319-7988
343
These gradients are important two- ϕt

point tensors providing transforma-
tions between objects related to the re- growth space
spective configurations. The commonly
used deformation gradient F is related
to the reference and actual configura-
tion. Thus, Lagrangian mapping from ρ0g , dvg
the reference configuration to the ac- Fg Fe
tual configuration ϕt can be described
reference config. actual config.
by the composition ϕt := µ ◦ κ−1 .
In analogy to growth in a living struc- F
ture we refer to a volumetric growth
approach. We apply a multiplicative ρ∗0 , ρ0 , dV γ ρt , dv
decomposition of the deformation gra-
dient F into an elastic part Fe and a K
T
M
growth part Fg as introduced in [4, 6, parameter space
8]. Following Lubarda and Hoger [7],
we apply an isotropic ansatz for the κ Zi µ
growth part of the deformation gradi-
ρΘ , dvΘ
ent with the stretch ratio ν
F = Fe Fg Figure 1. Kinematic concept
(1)
Fg = ν 1X
b.
Mass exchange is realized by a mass source/sink term R0 in the reference configuration which leads to
time dependent update of the initial mass density ρ∗0 to the resulting mass density ρ0 , i.e.
Z t
∗
ρ0 = ρ0 ± R0 dt. (2)
t0
1.2 Constitutive Framework
As postulated in the second law of thermodynamics, a system’s entropy never decreases. On this basis,
we shortly present a thermodynamically consistent evaluation of the coupled chemical and mechanical
problem motivated by the outline of Hu and Shen in [5]. The evaluation yields the mechanical stresses
and the description of the chemical diffusion. We apply the fundamental Gibbs equation for reversible
processes with chemical impact as follows
ρt 0 = θ s0 + T : D + µγ c0γ , (3)
wherein is the internal energy, θ is the temperature, s is the entropy, T is the Cauchy stress tensor
in the actual configuration, D is the velocity gradient, µγ are the chemical potentials and cγ are the
concentrations. With the local form of the balance equation of energy, the time derivative of the total
entropy results in
1 R0
s0 = [ µγ cγ div x0γ − ]. (4)
θ J
Herein, x0γ is the velocity vector of the concentrations evaluated by the balance of mass of the concen-
trations with c0γ = − cγ div x0γ . As presented in Hu and Shen [5], the internal dissipation energy in the
medium is transformed to the dissipation heat and produces the irreversible entropy. With this, the total
entropy can be decomposed in a reversible and in an irreversible part s0 = s0(r) + s0(i) . We evaluate the
reversible part of the entropy
R0 µγ cγ x0γ
s0(r) = − + div[ ], (5)
θJ θ
wherein the irreversible part of the entropy can be calculated by s0(i) = s0 − s0(r) , i. e.
1
s0(i) = [− x0γ · grad (cγ µγ )]. (6)
θ
344
We use the Legendre transformation to obtain the Helmholtz function ψ = − θ s. With the evaluation
of the local energy balance and the entropy, we identify the time derivative of the Helmholtz function
under isothermal conditions, namely
R0 1 R0
ψ0 = − + (D : T) + + µγ c0γ + x0γ · grad (cγ µγ ). (7)
ρt J ρt J
Furthermore, we separate the elastic and dissipative part of the Helmholtz energy function as follows
0 1
elastic part : ψ el
:= (D : T) + µγ c0γ
ρt (8)
0
dissipative part : ψ dis
:= x0γ · grad( cγ µγ ).
We can decompose the ansatz for the Helmholtz function in a mechanical, applying the neo-Hookean
hyperelastic model ψ Neo , and a chemical part ψ Chem . To ensure the entropy inequality, we evaluate the
following equations for the stress T, the chemical potential µγ and the velocity of the concentrations x0γ
δψ Neo T
T = 2 ρt F F
δC
∂ψ Chem (9)
µγ =
∂cγ
x0γ = −D grad(cγ ),
wherein D is the diffusion coefficient.
1.3 Numerical Solution
The previously introduced material formulation for the coupled diffusion and growth/degradation model
leads to a reduction of the unknown quantities and to a set of coupled differential equations. A numerical
approximation, the FEM, is applied for further procedures. In order to close the system of partial
differential equations, the unknown quantities are solved by the choice of governing equations. This
approach is used to solve the set of degrees of freedom, including the displacement u and the concentration
cγ evaluated with the balance of momentum and the balance of mass for concentrations. The weak forms
together with T = 1J P FT results in
R R
P : Grad (δu) dV = t · δu dA
B0 ∂B0
R R (10)
c0γ δcγ − x0γ · grad ( δcγ ) dV = x0γ · n δcγ dA.
B0 ∂B0
345
2 Structural Optimisation
Figure 2 gives an overview of the optimisation algorithm. The

entire program runs in a Matlab framework, which includes
a connection to the open software gmsh for mesh creation on
the one hand, and an interface to a Fortran code using MEX-
file interfaces on the other. So, the model presented here is
embedded as a Fortran code using MEX-file interfaces. We
apply Matlab to access the integrated optimisation toolboxes.
Purpose of the optimisation problem is a computational de-
sign framework accompanied by von Mises stress and design
restrictions. In this example we focus on a simple example,
namely a structure with a hole. For the optimisation of the
initial problem, an objective function, design parameters and
Figure 2. Algorithmic concept side conditions are required. In this example, the objective
function J(s) is the area of the structure aiming a reduction
of material using the parameters of the hole as the design parameters.
3 Numerical Example
We consider a structure with a hole, which is attacked by chemical concentrations from the left side. The
impact of the concentrations, results in degradation of the material. Figure 3 shows the time depending
development of the degradation gradient Fg triggered by the concentrations. The optimisation algorithm
enables a design reduction, thus holding the side conditions.
Figure 3. Left: Boundary condition of the concentration and displacement. Middle: Contour plot, showing the
distribution of the concentration in the last time step and the impact on Fg in the marked point. Right:
Optimised design with a material reduction of 17.38%.
References
[1] F.-J. Barthold. Zur Kontinuumsmechanik inverser Geometrieprobleme. TU Braunschweig, 2002.
[2] F.-J. Barthold. Remarks on variational shape sensitivity analysis based on local coordinates. Engi-
neering Analysis with Boundary Elements 32 (2008), pp. 971–985.
[3] F.-J. Barthold, N. Gerzen, W. Kijanski, and D. Materna. Efficient Variational Design Sensitivity
Analysis. In: Mathematical Modeling and Optimization of Complex Structures. Springer International
Publishing, 2016, pp. 229–257.
[4] G. Himpel, E. Kuhl, A. Menzel, and P. Steinmann. Computational modelling of isotropic multiplica-
tive growth. Computer Modeling in Engineering & Sciences 8 (2005), pp. 119–134.
[5] S. Hu and S. Shen. Non-equilibrium thermodynamics and variational principles for fully coupled
thermal–mechanical–chemical processes. Acta Mechanica 224 (2013), pp. 2895–2910.
[6] D. Kuhl, F. Bangert, and G. Meschke. Coupled chemo-mechanical deterioration of cementitious
materials. Part I: Modeling. International Journal of Solids and Structures 41 (2004), pp. 15–40.
[7] V. A. Lubarda and A. Hoger. On the mechanics of solids with a growing mass. International Journal
of Solids and Structures 39 (2002), pp. 4627–4664.
[8] A. Menzel and E. Kuhl. Frontiers in growth and remodeling. Mechanics Research Communications
42 (2012), pp. 1–14.
346

Acoustic-structure interaction in the Scaled Boundary

Finite Element Method for prismatic geometries
1
Paul Wasmer ∗
, Jannis Bulling1 , Hauke Gravenkamp2 and Jens Prager1
1 Bundesanstalt für Materialforschung und -prüfung
2 University of Duisburg-Essen, Germany
Abstract. Due to the short wavelength compared to the dimensions of the structure, the simulation of ultra-
sonic waves is still a challenging task. A numerical method wellisuited for this purpose is the semi-analytical
Scaled Boundary Finite Element Method (SBFEM). When applying this method, only the boundary of a
computational domain is discretized using finite elements, while the interior is described by an analytical
ansatz. Hence, the number of degrees of freedom is reduced significantly compared to the classical Finite
Element Method (FEM). In recent years, a particular formulation of the SBFEM for the simulation of ul-
trasonic guided waves was developed. The method constitutes an efficient algorithm for prismatic struc-
tures of arbitrary length, such as plates, pipes, or beams. Wave propagation phenomena in such struc-
tures can be modeled for isotropic and anisotropic inhomogeneous waveguides. Even though the method
is an efficient tool for the simulation of guided waves in solid media, a reliable model for the simulation
of acoustic wave propagation in fluids as well as acoustic-structure interaction in terms of SBFEM is still
missing. In principle, the fluid can be described by a displacement-based formulation and thus be im-
plemented in existing SBFEM algorithms for solid bodies. However, due to the discretization with clas-
sical finite elements, spurious modes occur, which cannot be separated from the physical modes straight-
forwardly. The spurious modes can be suppressed using a penalty parameter. Although very accurate re-
sults were achieved for some problems, this procedure has been proven unreliable for certain cases. For
this reason, we propose a different approach in this contribution. We employ a pressure model to simu-
late the acoustic behavior of fluids. The implementation of the pressure model results in a higher effort
due to the necessity of incorporating coupling terms, but it presents a stable alternative without spurious
modes. The accuracy of the method is demonstrated in comparison with results obtained using the FEM.
Introduction
Prismatic geometries, e.g. plates or pipes, act as waveguides for ultrasonic acoustic waves. In waveguides
waves can travel long distances without decaying. The physical behavior in waveguides is very complex
as the ultrasonic waves show strongly dispersive behavior. An efficient numerical tool to simulate ultra-
sonic guided waves is therefore essential to investigate the physical behavior. Besides other numerical
methods, the Finite Element Method (FEM) can be used to simulate wave propagation. However, due
to the short wavelengths at ultrasonic frequencies, a very fine mesh is required. Thus, FE models result
in large numerical systems and hence, long computation times. An efficient alternative is given by the
Scaled Boundary Finite Element Method (SBFEM) [4]. In the SBFEM, a semi-analytical method, only
the boundary of the computational domain is discretized with finite elements and hence, the number of
degrees of freedom is reduced compared to the FEM. The physical behavior in the interior of a compu-
tational domain is described analytically. A special formulation of the SBFEM for prismatic geometries,
e.g. plates and pipes, was developed by Gravenkamp et al. [1, 2, 3].
The method can be used to simulate wave behavior in isotropic and anisotropic inhomogeneous solid
waveguides. However, a stable method for the simulation of acoustic behavior in fluids and the simula-
tion of acoustic-structure interaction is still missing. In a first attempt, a displacement-based formulation
was used to describe the fluid. Due to the discretization by finite elements, this approach results in
spurious modes, a phenomenon widely reported in the literature. The spurious modes can be suppressed
by applying a penalty parameter [5]. However, this method does not work reliably for all cases.
∗ Corresponding author: paul.wasmer@bam.de
347
(a) (b)
Figure 1. a) 2D domain discretized with the SBFEM; b) prismatic semi-infinite 2D geometry discretized with the
SBFEM
A pressure model constitute a more natural and straightforward approach to describe the acoustic be-
havior of a fluid. For the application of fluid-structure interaction, this method results in a higher imple-
mentation effort compared to the displacement-based approach, as coupling terms between the acoustic
and the structural domain are needed. However, to the best knowledge of the authors, no numerical
instabilities are reported in the literature. Thus, we employ the pressure model in the SBFEM for the
computation of dispersion curves. To show the accuracy of the method, results are compared to those
obtained using the finite element software COMSOL Multiphysics.
The SBFEM relies on a coordinate transformation. In general, the Cartesian coordinates are mapped
onto the so-called SBFEM coordinates, η and ξ. η is the local coordinate of the finite elements on the
boundary (see Fig. 1a). ξ is the scaling direction pointing from the so-called “scaling center” to the
boundary. In the case of prismatic geometries the scaling center is placed at x = −∞. It follows that ξ is
equivalent to the Cartesian x coordinate. It therefore only becomes necessary to discretize the boundary
along the y direction. By applying the coordinate transformation, the partial differential equation is
transformed into an ordinary differential equation. In the case of prismatic geometries, a linear ordinary
differential equation is obtained.
1.1 Acoustic-structure interaction
The algorithm to describe acoustic wave behavior in solids can be found in [1, 3] and is therefore not
repeated here. The acoustic behavior of a fluid using pressure as the primary variable is described by the
Helmholtz equation
K∆p − ρf p̈ = 0 , (1)
where K is the bulk modulus and ρf the density of the fluid. The divergence of the pressure in 2D is
given by
∇p = D1 pn ,x (x) + D2 pn (x) (2)
where pn (x) is a vector containing the nodal pressure values and D1 and D2 are given by
| 1 |
D1 = 1 0 Np D2 = 0 1 Np ,η . (3)
y,η
Np is a vector containing the shape functions. Through applying the coordinate transformation to
SBFEM coordinates and performing some computation steps (equivalent to the computation steps for
solid found in [1]), the SBFEM equation in the frequency domain is obtained
E0p pn ,xx (x) − E2p pn (x) − ω 2 M0p pn = 0 . (4)
E0p , E2p and M0p are coefficient matrices integrated over the discretized boundary
Z Z Z
E0p = D1 | KD1 |J|dΓ1 , E2p = D2 | KD2 |J|dΓ1 , M0p = Np | ρf Np |J|dΓ1 . (5)
348
The boundary condition between fluid and solid in the frequency domain is given by
∂p
K = ω 2 ρf n| u , (6)
∂n
where n is the normal vector pointing from the fluid domain into the solid domain. For the SBFEM, this
results in the following equation system

E0s 0 un,xx E1s | − E1s 0 un,x
+ − (7)
0 E0p pn ,xx 0 0 pn ,x

E2s Q un M0s 0 un
+ ω2 = 0, (8)
0 E2p pn −ρf KQ| M0p pn
where E0s , E1s , E2s and M0s are the coefficient matrices of the solid domain. A detailed description
of their computation can be found in [1]. The vector un contains the nodal displacements. In 2D the
coupling matrix is reduced to Z
Q= N| nNp dΓ2 = 1 (9)
where N are the shape functions of the solid element and Γ2 the discretized boundary between solid and
fluid. The dispersion behavior is obtained by solving the eigenvalue problem (8). We apply the algorithm
proposed in [1].
2 Examples
To demonstrate the accuracy of the method, two examples are presented. In the first example the
dispersion curves of two aluminium plates with a thin liquid layer in between are calculated (see Fig.
2a). Both plates have a thickness of h = 5 mm. The liquid layer in between has a height of d =
kg
1 mm. Both plates are made from aluminium (E = 70 GPa, ν = 0.33, ρ = 2700 m 3 ). As liquid water
kg
(K = 2.19 GPa and ρf = 1000 m3 ) is used. The dispersion behavior of the phase velocity is shown in
Fig. fig:twoPlatesResults. The results are in perfect agreement with dispersion curves computed with
COMSOL Multiphysics.
(a)
(b)
Figure 2. Dispersion behavior of plate with a fluid layer in between. a) simulated geometry b) phase velocity over
frequency.
In a second example, the dispersion behavior of a fluid-filled pipe is computed. The pipe is made of
kg
steel (ρ = 7850 m 3 , ν = 0.3 and E = 200 GPa) with an inner radius of r = 5 mm and a wall thickness
kg
of h = 1 mm (see Fig. 3a). The inside is filled with water (ρf = 1000 m 3 and K = 2.19 GPa). Here, an
axisymmetric model is used, where in the SBFEM only one radius of the pipes has to be discretized with
349
one-dimensional finite elements [2]. Hence, very short computation times are achieved. In Fig. 3b the
dispersion behavior is displayed. The results of SBFEM are compared to results from COMSOL Mul-
tiphysics. The results from both methods are in excellent agreement. By comparing the computational
times on the same machine (Intel Xeon X650, 96 GB RAM) for both methods the efficiency of SBFEM
becomes obvious: The implemted SBFEM algorithm takes 36 s, whereas FEM is much more expensive.
COMSOL Multiphysics runs 7 min 30 s to produce the shown results.
(a) (b)
References
[1] H. Gravenkamp, C. Song, and J. Prager. A numerical approach for the computation of dispersion
relations for plate structures using the scaled boundary finite element method. Journal of Sound and
Vibration 311 (11 2012), pp. 2543–2557.
[2] H. Gravenkamp, C. Birk, and C. Song. The computation of dispersion relations for axisymmetric
waveguides using the Scaled Boundary Finite Element Method. Ultrasonics 54 (2014), pp. 1373–
1385.
[3] H. Gravenkamp, C. Birk, and C. Song. Simulation of elastic guided waves interactiong with defects
in arbitrarily long structures using the scaled boundary finite element method. Journal of Compu-
tational Physics 295 (2015), pp. 438–455.
[4] C. Song. The scaled boundary finite element method in structural dynamics. International Jounral
for Numerical Methods in Engineering 77 (2009), pp. 1139–1171.
[5] P. Wasmer, F. Krome, J. Bulling, and J. Prager. Prismatic semi-analytical elements for the simulation
of linear elastic problems in structures with piecewise uniform cross section. PAMM 18 (2017),
e201800139.
350

Modelling the Soil-Water Interaction at Fluid-Inltrated

Roads under Loading of Steady State Rolling Tires
Ines Wollny
∗ 1, WaiChing Sun
2 and Michael Kaliske
1
1 Institute for Structural Analysis, Technische Universiät Dresden, 01062 Dresden, Germany
2 Department of Civil Engineering and Engineering Mechanics, Columbia University, 4709 New York, NY, USA
Abstract. In case of tires driving over a uid-inltrated subsoil like o-road on soils or on porous and perme-
able pavements, the interaction between the solid and the uid constituent inside the underground aects the
tire-subsoil interaction. This hydro-mechanical coupling eect is captured via poro-elasticity theory. In order
to achieve a numerically ecient description of the tire rolling over the uid-inltrated subsoil, the porome-
chanical balance equations are formulated in an arbitrary Lagrangian Eulerian (ALE) frame and account for
nite deformation. The solution of the poromechanical ALE model of the uid-inltrated subsoil via a nite
element (FE) formulation is performed by an operator split scheme. Thereby, dierent split assumptions
(drained, undrained, semi-drained) are investigated for the solid step. This staggered solid-uid iteration
is embedded into a coupled tire-subsoil interaction model, which is solved via a sequential iteration scheme
as well. The whole tire-soil-water interaction is, then, solved within a hierarchical operator split scheme.
1 Poromechanical ALE formulation
In case of uid-inltrated porous media, uid constituents (e.g. water and air) are present in addition to
the solid constituent. In the poromechanics, these phases are not treated separately but as multi-phase
continua. A fully saturated soil is considered as two-phase (solid-water) continuum while an unsaturated
soil is modelled as a three-phase continuum (solid-water-air) with the corresponding volume fractions
φα = vα /v of the constituents α = s, w, a, which
P is dened by the current volume vα of the constituent
in proportion to the current volume v = α α of the total mixture. In contrast to a single-phase
v
continua that undergoes compatible deformation, the uid and solid constituents of a uid-inltrated
porous medium may move at dierent trajectories, see Fig. 1. Thus, a controlled mass of solid constituents
may contain uid in the deformed void space that was not necessarily present in the void space of the
solid skeleton in the reference conguration.
Figure 1. Poromechanical ALE kinematics for a two-phase continuum
∗ Corresponding author: ines.wollny@tu-dresden.de
351
The solid and the uid constituents occupy the combined space Φ(B) in the current conguration. In
the initial conguration B contains the solid phase and B f contains the pore uid. For the arbitrary
Lagrangian Eulerian (ALE) formulation, an additional reference conguration is introduced, where the
solid phase is in space χ(B) and the uid phase is described by χ(B f ). Thereby, the mapping from the
initial to the reference conguration is a rigid body motion. Then, the current position x of a point P
at time t of combined solid and uid can be expressed in terms of the initial solid position X or in terms
of the initial uid position X f by

x = Φ (X, t) = Φ̂ (χ (X, t)) = Φf X f , t = Φ̂f χ X f , t . (1)
By introducing, further, a moving reference coordinate system eALE

i in the ALE frame, the current
position of a point P might be expressed in several ways by
x = X + u = X f + uf = χALE
 + ϕ = χALE
 + χ + û = χALE
 + χf + ûf . (2)
In the treated case of uid-inltrated roads under rolling tire load, the introduced reference frame moves
with the tire through the space. In case of steady state rolling tires and a subsoil, which is homogeneous in
driving direction, the deformation state of the tire and the subsoil with respect to the introduced moving
reference frame becomes time-independent. This feature enables a numerically ecient description of
the tire-subsoil interaction. It is worth noting that treating the tire-soil interaction with respect to the
introduced moving coordinate system leads to a ow of the material particles through the reference
conguration. The deformation gradient can be expressed as
∂x ∂ϕ ∂ϕ ∂χ
F = = = · = F̂ · R, (3)
∂X ∂X ∂χ ∂X
where F̂ is the deformation gradient with respect to the reference conguration and R contains the rigid
body rotations from the initial to the reference conguration with det R = 1.
The total Cauchy
P stress of the mixture is obtained by summing up all partial stress tensors of the con-
stituents σ = α σ α . Postulating that only volumetric stresses act on the water and the air constituent,
their partial stress tensors read
σ α = −φα pα I, α = w, a (4)
with the macroscopic pore pressure pα . The eective stress σ 0 that is energy-conjugate to the rate of
deformation of the solid matrix is obtained by
σ 0 = σ + B (Sw pw + (1 − Sw )pa ) I (5)
with the Biot's coecient B and the degree of saturation Sw = φw /(1 − φs ). Please note that the fully
saturated case is obtained for Sw = 1. Subsequently, it is assumed that the air pressure is negligible
pa = 0.
The continuum mechanical description and the corresponding balance principles for the multi-phase
continua are derived based on the solid trajectory and the solid deformation mapping with respect to the
reference conguration. The balance of linear momentum, which is the base for the solid step in the later
described staggered solution, with respect to the reference conguration reads
ρ̂ v̇ = Div P̂ + ρ̂ b. (6)
Thereby, ρ̂ is the density in the reference frame, volume loads are represented by b, the divergence is
Div A = ∂Aab /∂χb and v̇ is the time derivative of the solid material velocity v (x, t). The rst Piola-
Kirchho stress tensor with respect to to reference frame is obtained from P̂ = Jσ · F̂ −T .
The balance of the pore water mass content in the solid skeleton, which is the base for the uid step in
the later described staggered solution, with respect to the reference frame is
Ṁ w + Div ω̂ = 0. (7)
−1
Thereby, M w = Jφw ρw is the water mass content and ω̂ = J F̂ · ω is the relative mass ow with respect
to the reference frame based on the Eulerian relative mass ow ω = φw ρw (v w − v), which is assumed to
be Darcian. The corresponding nite element (FE) equations of the balance equations are obtained by
applying Galerkin's method, including the constitutive relations, linearization and discretization, compare
[2] for the fully saturated case.
352
2 Sequential solid-uid solver
The mass balance equation as well as the balance of linear momentum depend both on the current
deformation state ϕ and the pore water pressure pw . To solve this coupled system of equations, a
staggered solution procedure is applied.
The staggered solid-uid-iteration starts with a solid step by evaluating the balance of linear momentum
(Eq. (6)) to update the current deformation state ϕ. Thereby three dierent split assumptions are studied
for the solid step. In case of a drained split, the uid mass content might change (drainage) while the
pore water second Piola-Kirchho stress tensor Ŝ w = −JBSw pw F̂ −1 · F̂ −T remains constant during the
solid step. At an undrained split, it is assumed that the pore water mass content does not change during
the solid step. This condition is obtained by prescribing Div ω̂ = 0, which nally leads to a prediction
of the updated pore water pressure p? . The drained split assumption neglects the stiness stemming
from the presence of the pore water during the solid step and leads, thus, to an overestimation of the
solid deformations that might cause oscillating or non-converging solid-uid-iterations. In contrast, the
undrained split assumption overestimates the stiness stemming from the presence of the pore water
during the solid step and leads, thus, to an underestimation of the solid deformations that might cause
slowly converging solid-uid-iterations. Therefore, a semi-drained split is studied as third case that is
obtained by a modication of the undrained pore water predictor.
After the solid step, the pore water pressure pw is updated by evaluating the mass balance equation
(Eq. (7)). During the uid step, the solid deformation is xed. If the residual of the balance of linear
momentum using the updated pore water pressure pw is smaller than a prescribed tolerance, the solid-
uid-iteration is terminated. Otherwise, it restarts with a new solid step.
3 Coupled tire-soil-water interaction
To enable a realistic representation of tires rolling over a fully/partially saturated soft subground (soil), the
developed poromechanical ALE formulation is implemented into the coupled tire-pavement computation
scheme presented in [1]. Since, the tire and the pavement are simulated in dierent FE codes, this tire-
pavement computation is organized by a program interface as staggered iteration between the FE ALE
tire model and the FE ALE pavement model. The implementation of the staggered solid-uid iteration
scheme for the poromechanical ALE formulation into the staggered tire-pavement computation scheme
demands for a hierarchical solution strategy, see Fig. 2, where the tire-pavement or rather the tire-soil
iteration is the outer iteration loop, while the solid-uid interaction is solved during an inner iteration
loop at the pavement or soil computation.
Figure 2. Hierarchical sequential coupling procedure for tire-soil interaction
This staggered tire-soil iteration scheme is applied, subsequently, for the computation of a hyperelastic
truck tire loaded by 2 tons (=19.62
ˆ kN) that is rolling over a fully saturated soil subground at dierent
driving velocities. The solid soil material is represented by the hyperelastic Neo-Hookean model. The
applied material parameters of the fully saturated soil are given in Tab. 1. As boundary conditions
for the discretized FE model of the soil segment, all displacements at the bottom as well as the normal
displacements at the lateral surfaces are xed. Further, the pore pressure is zero at the whole top surface,
while all other boundaries are undrained.
353
Table 1. Material parameters of fully saturated soil

2 2 2
solid µ [kN/m ] K [kN/m ] Ks [kN/m ] B [−]
22 500 40 000 4 · 10 8
1
uid k [m /kNs]
4 w
φ̂ [−] Kw [kN/m ]
2
3.06 · 10−6 0.3 80 000
The resulting distribution of the pore water pressure in the soil is shown in Figure 3 for two dierent
speeds. As expected, the pore pressures are higher at larger speed, since the time for drainage is shorter
in this case. Fig. 4 illustrates, additionally, the resulting surface displacements of the saturated soil along
the driving lane for the two dierent speeds as well as for the interaction of the tire with a dry soil
(without pw ). The displacements are largest in the dry case, since the solid phase has to carry the whole
tire load. In case of the fully saturated soil, the pore water pressure increases at increasing speed. Thus,
the solid eective stress and the solid displacements decrease at increasing speed.
Figure 3. Pore water pressure in plane along driving lane for guiding velocity of a) w1 = 0.5 m/s, b) w1 = 1 m/s
Figure 4. Vertical displacements of the surface along the driving lane for the three computation cases
Acknowledgments
The nancial support of the German Research Foundation (DFG) of Research Unit FOR 2089 under
grant KA 1163/30 is gratefully acknowledged.
References
[1] M. Kaliske, I. Wollny, R. Behnke, and C. Zopf. Holistic Analysis of the Coupled Vehicle-Tire-
Pavement System for the Design of Durable Pavements. Tire Science and Technology 43 (2015),
pp. 86116.
[2] I. Wollny, W. Sun, and M. Kaliske. A hierarchical sequential ALE poromechanics model for tire-
soil-water interaction on uid-inltrated roads. International Journal for Numerical Methods in En-
gineering 112 (2017), pp. 909938.
354
MS 18: Crack initiation and crack growth in modern composite materials: fracture of interfaces, material
discontinuities and composite laminates

Constitutive modelling of carbon fibre reinforced poly-

mers for adhesively bonded cfrp-steel joints
Michael Donhauser1 and Anton Matzenmiller∗,1
1 Institute of Mechanics, Dept. of Mechanical Engineering, University of Kassel, Mönchebergstr. 7, 34125 Kassel, Germany
Abstract. This contribution comprises the constitutive modelling and finite element failure analysis of ad-
hesively bonded cfrp-steel joints under quasi-static loading with the focus on the modelling of the cfrp. A
transversely isotropic material model is utilised for the elastic-brittle behaviour of the cfrp, where a nonlinear
ansatz takes the axial shear behaviour into account. The inter-fibre failure and postcritical behaviour of the
cfrp are modelled by Puck’s failure criterion and a failure mode based anisotropic damage approach. Delam-
ination is considered by means of an interface model with a bilinear traction-separation-law. The adhesive
layer is characterised by the so-called Toughened-Adhesive-Polymer (TAPO) model, for elasto-plasticity and
isotropic damage. An elasto-plastic material model is used for the steel components. For four of the six test
cases, the numerical validation results under quasi-static loading show a good agreement with the test data.
1 Introduction
Today, material compounds with enhanced mechanical properties are made of carbon fibre reinforced
polymers (cfrp) and steel components by means of adhesive joining. In order to optimise the structural
strength of the material compound with the finite element method, material models must be used that
capture the elastic and failure behaviour of each individual material accurately. Especially, the failure
modelling of cfrp is a challenging task due to the complex failure mechanisms and the interaction of
different failure modes causing ultimate fracture. In [6], material models, which are implemented in the
commercial FE-software LS-DYNA, are used to perform FE-simulations for cfrp and adhesive-cfrp-steel
joints that exhibit a simple stress state. The lack of capturing the nonlinear axial shear behaviour as
well as the postcritical behaviour after a first ply failure, leads to improvable results. Therefore, in
[7], a detailed model for the cfrp is developed, which captures the nonlinear axial shear behaviour, the
intralaminar failure due to fibre and inter-fibre breakage, the postcritical behaviour after a first ply failure
and the interlaminar failure i. e. delamination. The constitutive model of the cfrp is presented in this
contribution.
2 Constitutive Modelling
2.1 Carbon Fibre Reinforced Polymer

The elastic material behaviour of the unidirectional layer (ud-layer), depicted in Fig. 1, is described by
the transverse isotropic model. The model is formulated in the material coordinate system (x1 , x2 , x3 ),
where x1 denotes the fibre direction. For the normal strain the strain-stress relation reads as
( )  1 ( )
ε11 E1 − νE121 − νE121 σ11
ε22 =  1
E2 − νE232  σ22 , (1)
ε33 sym. 1 σ33
E2
whereby E1 is the Young’s modulus in fibre direction, E2 perpendicular to the fibre direction and ν12 is
the Poisson’s ratio in the coordinate direction x1 x2 . The transverse shear behaviour is described with
the transverse shear modulus G23 by the linear ansatz:
E2
τ23 = G23 γ23 with G23 = , (2)
2(1 + ν23 )
∗ Corresponding author: e-mail post-structure@uni-kassel.de, phone +49 (0)561 804 2043, fax +49 (0)561 804 2720
355
and the axial shear behaviour by the nonlinear ansatz with the model parameters c0 -c3 :
τ1j (γ1j ) = c0 tanh (c1 γ1j ) + c2 tanh (c3 γ1j ) j = 2, 3 . (3)
Failure model The intralaminar failure of the cfrp is divided into inter-fibre and fibre failure. In Puck’s
exposure criterion, the normal stress σn and shear stress τnt and τn1 acting on the potential fracture plane,
which is inclined about the fracture angle θfp (see Fig. 1), are used to describe the inter-fibre failure in
an ud-layer. For tensile loading (σn ≥ 0) the exposure criterion reads as
v !2 !2
u 2
u p+ p+
+ t 1 ⊥ψ (θfp ) τnt (θ fp ) τn1 (θ fp ) ⊥ψ (θfp )
fE (θfp ) = +A
− A 2
σn (θfp ) + A
+ A
+ A σn (θfp ) = 1 (4)
R⊥ R⊥ψ (θfp ) R⊥⊥ R⊥k R⊥ψ (θfp )
and for compressive loading (σn < 0) as

v !2 !2
u
u τ (θ ) 2 p− p−
− t nt fp τn1 (θfp ) ⊥ψ (θfp ) ⊥ψ (θfp )
fE (θfp ) = A
+ A
+ A (θ )
σn (θfp ) + A (θ )
σn (θfp ) = 1 . (5)
R⊥⊥ R⊥k R⊥ψ fp R⊥ψ fp
In Eqs. (4)-(5) all quantities RA are fracture resistances acting on the action plane A and all p± are slope
parameters of failure envelopes, see [8] for details. The fracture stresses are functions of the global stress
state and the fracture angle θfp :
σn (θfp ) = c2 σ22 + s2 σ33 + 2csτ23 , τnt (θfp ) = −scσ22 + scσ33 + (c2 − s2 )τ23 , τn1 (θfp ) = cτ12 + sτ13 (6)
with c = cos(θfp ) and s = sin(θfp ). The fibre stress σ11 does not influence the inter-fibre failure and is there-
fore neglected in Eq. (6). In order to evaluate the exposure criterion, it is always necessary to determine
the fracture angle θfp for a given stress state (σ22 , σ33 , τ12 , τ23 , τ13 ). The fracture angle is defined as that
one, for which the exposure criterion reaches the maximum value. Hence, for tension (“+“) or compression
(“-“) we solve the corresponding nonlinear scalar equation x3 x3
xt
σ 33
∂fE± (θfp ) τnt xn
=0 with θfp ∈ [−90◦ , 90◦ ] (7) τ31
∂θfp τ32 τ23 σn
τ13
using Newton’s method. The obtained fracture angle from τ12 τ
σ22 τn1 θfp
σ x1 x2
(7) and corresponding stress state from (6) are used to 11 21
x2 x1
x1
evaluate the exposure criterion (4) or (5), while an inter- fracture plane
fibre failure occurs if fE± ≥ 1.
The fibre failure under tensile loading is taken into ac- Figure 1. Ud-layer with stress state on fracture
plane
count by the maximum stress criterion
σ11
ffb = = 1 with Rk : tensile fibre strength . (8)
Rk
Post-critical behaviour In a multidirectional laminate, first ply failure due to an inter-fibre breakage
does not lead to abrupt loss of load-bearing. The load-bearing capacity is still remaining, but with a
reduced laminate stiffness. In this contribution, a failure mode based anisotropic damage approach with
six damage variables is applied to capture the stiffness reduction according to [4]. A 3×3 damage operator
with the damage variables ω11 , ω22 and ω33 arranged along the major diagonal, is introduced and applied
to Eq. (1) in order to obtain the compliance relationship in terms of the damage variables:
 
( ) 1
− νE121 − νE131 ( )
ε11 (1−ω11 )E1 σ11
ε22 =

1
(1−ω22 )E2 − νE232 
 σ22 . (9)
ε33 sym. 1 σ33
(1−ω33 )E3
The proposed method of [5] is applied to the components of the minor diagonal in Eq. (9) to retain the
symmetry of the compliance matrix. In case of damage, the linear and nonlinear ansatz of the shear
behaviour in Eq. (2) and (3) reads as:
τ23 = (1 − ω23 )G23 γ23 , τ1j = (1 − ω1j )(c0 tanh (c1 γ1j ) + c2 tanh (c3 γ1j )) j = 2, 3 . (10)
The Eqs. (4),(5) and (8) are used to define the loading functions
f2± = fE± − r2± = 0 and f1 = ffb − r1 = 0 (11)
356
with the variable thresholds

h i h i
r2± = max 1, max fE± and r1 = max 1, max (ffb ) (12)
−∞<τ <t −∞<τ <t
due to inter-fibre and fibre failure, respectively. For f2± < 0 and f1 < 0 there is no damage evolution,
thus r2± = 1 and r1 = 1. With the damage growth functions for the corresponding failure mode

± 1 − (r2± )m2 1 − (r1 )m1
φ2 = 1 − exp and φ1 = 1 − exp (13)
m2 m1
the damage evolution is defined as
ω = φ1 l + φ2 q (14)
T T T
with ω = [ω11 , ω22 , ω33 , ω12 , ω23 , ω13 ] , l = [1, l22 , l33 , l12 , l23 , l13 ] and q = [0, q22 , q33 , q12 , q23 , q13 ] . In
Eq. (13), the quantities m2 and m1 denote the strain softening parameters. The components of l and
q in Eq. (14) are parameters, which enable the coupling of the individual damage variables. In brittle
multidirectional laminates, which are considered in this contribution, abrupt and ultimate failure occurs,
when the fibre failure criterion in Eq. (8) is reached, thus φ1 in Eq. (14) is zero. The presented
constitutive equations for the cfrp are implemented as a user defined material model in LS-DYNA.
Delamination In order to describe the delamination process, interface elements are used in combination
with the interface model Cohesive-Mixed-Mode (*Mat_138) from LS-DYNA, which includes a bilinear
traction-separation-law and a quadratic mixed mode criterion according to [3]. The model parameters
are identified from test data of fracture mechanical DCB- and ENF-specimens.
2.2 Adhesive Layer and Steel Components

The adhesive layer is characterised by the so-called TAPO-model, which is an elasto-plastic model with
damage. More details about this material model can be found in [1] and [2].
The material behaviour of the steel is described by the simple model Piecewise-Linear-Plasticity (*Mat_24)
from LS-DYNA together with stress-strain test data of the steel at hand. The element deletion method
is applied, when the equivalent strain reaches the defined failure value.
3 Validation
The FE-simulation and failure analysis is carried out for the so-called thin shear specimen, depicted in
Fig. 2. The specimen consists of a steel adherend and a quasi-isotropic laminate, which are bonded
together by the adhesive layer. The quasi-isotropic laminate has 13 plies and an overall thickness of
2 mm. The thin shear specimen is tested for three adhesive thicknesses (dk =0.3, 1 and 2 mm) and
two laminate orientation (0◦ and 90◦ ) in order to investigate their influence on the failure behaviour.
For the laminate orientation of 0◦ the layup configuration adhesive layer qs laminate
steel
is [0◦ ,90◦ ,+45◦ ,-45◦ ,0◦ ,90◦ ,+45◦ ,90◦ , 0◦ ,-45◦ ,+45◦ ,90◦ ,0◦ ],
whereby the fibre direction of the top layer, which is at- vx
tached to the adhesive, coincides with the loading direction, ez
vx
see Fig 2. In the case of the 90◦ orientation, the fibres in the ey ex
top layer are arranged perpendicular to the loading direction
leading to the layup configuration [90◦ ,0◦ ,-45◦ ,+45◦ ,90◦ ,0◦ ,-
45◦ ,0◦ ,90◦ ,+45◦ ,-45◦ ,0◦ ,90◦ ]. Each of the thirteen single lay-
ers is discretised by a three-dimensional solid element in the orientation 0◦ orientation 90◦
FE-model, where interface elements are placed between ad-
jacent layers to take delamination into consideration. In the Figure 2. Sketch of thin shear specimen with
different laminate orientations
FE-simulation, failure can occur in the adhesive layer, the
steel component or in the laminate in form of a fibre breakage, an inter-fibre fracture or a delamination.
The material parameters are identified from test data of individual specimens of each material. The iden-
tification procedure and all identified parameters are documented in [7] in detail. In Fig. 3, the numerical
and experimental results are shown for a quasi-static loading (v = 1 mm/min). In four of six test cases
the numerically obtained stress-strain-curves correspond well with the experiments, see Fig. 3 a, b, d, f.
The structural strength is slightly and the failure strain strongly underestimated for the both other cases
(Fig. 3 c, e). For the laminate orientation of 90◦ in the test and in the simulation, failure occurs always
357
in the laminate due to delamination between the first and second ply caused by an preceding inter-fibre
breakage in the top layer, see Fig 3 for instance. In the test, failure in the adhesive layer occurs for the
laminate orientation of 0◦ and the adhesive thickness of dk = 0.3 and 1 mm. The FE-calculation predicts
the correct failure mode for the thinner adhesive layer. The test results of the specimen with dk = 2 mm
and laminate orientation 0◦ shows mixed failure in the adhesive layer and laminate, where the simulation
predicts only failure in the laminate. The validation results under quasi-static loading show that with
the presented method four of the six force-displacement-curves with the corresponding failure mode are
reproduced well. In both other cases (Fig. 3 c, e), the deviation in the force-displacement-curves is caused
by premature delamination in the FE-simulation, which will be part of forthcoming investigations.
12 12 8 8
Exp. failure adhesive Exp. failure laminate Exp. failure adhesive Exp. failure laminate
Sim. first inter-fibre break Sim. first inter-fibre break Sim. first inter-fibre break Sim. first inter-fibre break
nominal force [kN]
Sim. failure adhesive Sim. delamination Sim. delamination Sim. delamination

a) dk = 0.3 mm
6 c) dk = 1 mm 6
b) dk = 0.3 mm d) dk = 1 mm
8 orientation 0◦ 8 orientation 90◦ orientation 0◦ orientation 90◦
4 4
4 4
2 2
0 0 0 0
0 0.3 0.6 0.9 0 0.1 0.2 0.3 0.4 0 0.4 0.8 1.2 0 0.2 0.4 0.6
relative displacement [mm] relative displacement [mm] relative displacement [mm] relative displacement [mm]
8 8 Fracture pattern from [6]
Exp. mixed fracture Exp. failure laminate
Sim. first inter-fibre break Sim. first inter-fibre break Configuration
nominal force [kN]
Sim. delamination Sim. delamination 90◦ orientation

6 e) dk = 2 mm 6 dk = 1 mm
f) dk = 2 mm
orientation 0◦
orientation 90◦
4 4
Failure mode in experiment:
2 2 Initial inter-fibre breakage in first ply
and then delamination between 1st
and 2nd ply
0 0 → Same failure mode predicted in FE-simulation
0 0.4 0.8 1.2 0 0.2 0.4 0.6
relative displacement [mm] relative displacement [mm]
Figure 3. Validation results of adhesively bonded thin shear specimen with steel adherend and quasi-isotropic
laminate
References
[1] W. Böhme, M. Brede, F. Burbulla, C. Fehrenbach, O. Hesebeck, S. Kilchert, J. Lienhard, R.
Mahnken, S. Marzi, A. Matzenmiller, Michael May, D. Memhard, N. Nörenberg, L. Reissig, and H.
Voß. Robustheit und Zuverlässigkeit der Berechnungsmethoden von Klebverbindungen mit hochfesten
Stahlblechen unter Crashbedingungen. Abschlussbericht, FOSTA-Forschungsvereinigung. 2013.
[2] M. Brede, F. J. Heise, M. Schlimmer, A. Matzenmiller, R. Mahnken, O. Hahn, K. Dilger, P. Gumbsch,
K. Thoma, and O. D. Hennemann. Methodenentwicklung zur Berechnung von höherfesten Stahlkleb-
verbindungen des Fahrzeugbaus unter Crashbelastung. Abschlussbericht, FOSTA-Forschungsvereinigung.
2008.
[3] Pedro P Camanho and Carlos G Dávila. Mixed-mode decohesion finite elements for the simulation
of delamination in composite materials (2002).
[4] M. Chatiri and A. Matzenmiller. A Damage-Mode Based Three Dimensional Constitutive Model for
Fibre-Reinforced Composites. CMC: Computers, Materials & Continua 35.3 (2013), pp. 255–283.
[5] A. Matzenmiller, J. Lubliner, and R. L. Taylor. A constitutive model for anisotropic damage in
fiber-composites. Mechanics of materials 20.2 (1995), pp. 125–152.
[6] G. Meschut, M. Wünsche, A Matzenmiller, and M. Schmerbauch. Beanspruchungs- und fertigungs-
gerechtes Kleben von Faserverbundkunststoffen im Multi-Material-Design. Abschlussbericht, Fos-
chungsvereinigung Stahlanwendung FOSTA. 2015.
[7] G. Meschut, K. Henkel, A Matzenmiller, and M. Donhauser. Experimentelle und numerische Un-
tersuchungen zum Versagensverhalten von kalt ausgehärteten Stahl-FVK-Klebverbindungen unter
schlagartiger Belastung. Abschlussbericht, Forschungsvereinigung DECHEMA. 2018.
[8] A. Puck. Festigkeitsanalyse von Faser-Matrix-Laminaten: Modelle für die Praxis. Hanser, 1996.
358

Crack patterns at bi-material joints: An asymptotic

coupled stress and energy approach
Julian Felger ∗1
and Wifried Becker1
1 Technische Universität Darmstadt
Abstract. The effective strength of bi-material joints depends significantly on the elastic material properties
of the constituents as well as on geometrical parameters such as the orientation of the material interface. In
this work, a scarf joint specimen made of a polymer foam and aluminum is investigated using a four-point
bending specimen test with variable scarf angle. The critical load at failure is determined employing a coupled
stress and energy criterion relying on the concept of finite fracture mechanics where joint failure is defined
as spontaneous crack formation. Since a variety of potential crack configurations is admissible, an efficient
asymptotic model is employed reducing numerical effort significantly. Failure predictions according to the
asymptotic approach are compared to numerical reference data and experimental findings from literature.
Introduction
The use of lightweight design concepts employing composite materials such as sandwich structures and
adhesively bonded joints are frequently used in modern engineering structures. The material mismatch,
however, induces highly localised stress concentrations triggering crack initiation and subsequent catas-
trophic failure of the structure. Hence, a detailed analysis of such stress raisers combined with a physically
sound failure criterion is required to reliably assess the structural integrity. Stress-based approaches eval-
uate a stress criterion in a certain distance from the stress concentration [23] whereas classical fracture
mechanics approaches presume the existence of an inherent flaw [20]. These criteria, however, rely on
an empirical length which depends on material as well as on geometrical parameters of the structure. In
case of sharp notches or multi-material junctions failure criteria based on a critical value of the gener-
alised stress intensity factor (GSIF) have been applied [13]. This approach displays the drawback that
the critical GSIF does not represent a material property but depends on the structure’s geometry, for
instance the notch opening-angle [2] or the interface orientation in scarf joints [12].
In the present work, crack nucleation in a bi-material scarf joint is investigated employing a coupled stress
and energy criterion proposed by Leguillon [9]. The coupled criterion states that an energy as well as a
stress criterion must be fulfilled simultaneously forming a necessary and sufficient condition for fracture.
Following the concept of finite fracture mechanics for brittle failure as introduced by Hashin [7], the
process of crack initiation is modelled as an instantaneous formation of a finite sized crack. In contrast
to classical linear-elastic fracture mechanics, the finite crack length ∆a enters the problem as a new
unknown. Employing the coupled criterion allows for calculating the a-priori unknown finite crack length
∆a as well as the load at crack onset. Only material strength and fracture toughness as fundamental
material quantities enter the coupled criterion and no empirical length parameter is required. The coupled
criterion has been applied for strength predictions of numerous situations such as sharp V-notches [24],
rounded notches [11], open-hole plates [3, 4, 16, 22], adhesively bonded joints [6, 18] and has been used
to predict the nucleation of anticracks to model snow slab avalanche [15]. A review of finite fracture
mechanics can be found in [21]. In the present case of a bi-material scarf joint, multiple competing crack
configurations exist. Since all admissible crack configurations have to be considered, a high numerical
effort is needed in order to evaluate the coupled criterion. Thus, an asymptotic approach is followed
combining the method of complex potentials with the method of matched asymptotic expansions [8, 10].
This ultimately forms a semi-analytical framework allowing for efficiently model crack formation triggered
by the singularity at the bi-material notch. The resulting predictions of failure load, location of crack
∗ Corresponding author: felger@fsm.tu-darmstadt.de ORCID iD: ORCID (optional)
359
F/2 F/2
l0 l0
(b)
Aluminium Rohacell θ
x2
(a) h
ψ x1
O (c) ∆a
l/2 l/2 F/2
F/2
Figure 1. Bi-material scarf joint under four-point bending loading with an aluminum adherend on the left hand side
and a Rohacell adherend on the right hand side as used in [14]. (a) Uncracked scarf joint configuration
with interface angle ψ, height h = 40mm, thickness t = 40mm and lengths l0 = 150mm, l = 445mm.
(b) Notch induced failure mode. (c) Bending failure mode.
onset and angle of crack initiation are discussed in detail and compared to experimental findings from
literature [14].
1 Failure model: the coupled stress and energy criterion
Consider a bi-material scarf joint under four-point bending loading as depicted in Fig. 1(a). The joint
consists of a Rohacell 51 foam (material data according to [1]) and an aluminium adherend (EA = 70
GPa, νA = 0.3). Only cohesive failure of Rohacell is considered. The elastic contrast between both
adherends induces mathematical singular stresses at the bi-material point O. Hence, it is assumed that
cracks might emanate from O as show in Fig. 1(b). However, the constraint imposed by a small interface
angle may also lead to crack initiation at a certain distance from the bi-material point as depicted in
Fig. 1(c). The situation which ultimately occurs is related to the smallest failure load Ff and will be
determined by the underlying failure model. The coupled criterion requires a stress criterion to be fulfilled
along the potential crack path ΓC . As stress criterion, a quadratic interaction law of the form
s 2 2
σnn τnt
f= + ≥ 1, along ΓC , (1)
σc τc
is employed accounting for mixed-mode conditions. In addition, an energetic criterion must be satisfied
for crack onset to occur. The energy criterion is based on the energy balance and may be written as
∆Π
− = Ḡ ≥ Gc . (2)
∆a
Here, Ḡ is denoted as incremental energy release rate. The resulting failure load Ff is defined as the
minimum load F such that both criteria are fulfilled simultaneously:

Ff = min F f (σij ) ≥ 1 along ΓC and Ḡ ≥ Gc . (3)
∆a,θ,F
Solving Eq. (3) finally yields the failure load Ff , the crack length ∆a and the crack initiation angle θ.
2 Asymptotic expansion of stresses and incremental energy release rate
The stress field in the vicinity of the notch-tip can be expressed in form of a coordinate expansion with
respect the distance r:
σ(r, φ) = K1 rλ1 −1 σ 1 (φ) + K2 rλ2 −1 σ 2 (φ) + K 2 rλ2 −1 σ 2 (ϕ) + o(rRe[λ2 −1] ), as r → 0. (4)
Here, λi are characteristic exponents defining the behaviour of the stresses in radial direction, σ i (φ)
are eigenfunctions depending only on the circumferential coordinate, Ki represent the generalised stress
intensity factors (GSIFs) and (·) is the complex conjugate of a quantity. The exponents λi as well as the
eigenfunctions are determined analytically using the complex potential method and GSIFs are calculated
by means of a path independent integral, see [5, 10, 17] for more details. Assuming the crack length
360
(b)
(a) ψ =50° ψ =70° ψ =90°
/2
/2 /2
/2
ψ
ψθ ψ =110° ψ =130°
(c)
300 150
[N] −→
Crack angle θ [◦ ] −→
bending notch
bending
200 asym. 100
f
F numeric numeric
ce
Failure load
erfa asym.
100 50 int
experiment
0 0
60 80 100 120 140 60 80 100 120 140
Interface inclination ψ [◦ ] −→ Interface inclination ψ [◦ ] −→
Figure 2. Failure of Aluminum-Rohacell scarfed joints with respect the scarf angle ψ. (a) Comparison between
predicted failure loads Ff according to the asymptotic approach and numerical reference data as well
as experimental data. (b) Schematic of experimentally observed crack paths according to [14]. (c)
Predicted the crack initiation angles θ.
∆a to be small compared to a characteristic structural length, that is the beams’ height h, the crack
length can be understood as a small perturbation parameter. Using the method of matched asymptotic
expansions allows for deriving a series expansion of the incremental energy release rate for small values
of crack length, cf. [5, 8]:

Ḡ = K12 ∆a2λ1 −1 A11 + 4Re K1 K2 A12 ∆aλ1 +λ2 −1 + o ∆aλ1 +λ2 −1 , as ∆a → 0. (5)
The numerical scaling coefficients Aij depend on the interface angle ψ and are computed numerically, cf.
[5]. Note that only two finite element analyses for each interface angle are necessary to determine the
quantities K1 , K2 , A11 , A21 . In contrast, following a purely numerical approach numerous finite element
analyses of the cracked configuration must be carried out to calculate the incremental energy release in
terms of the unknown crack length. The derived expansions in Eq. (4) and Eq. (5) serve as a basis to
efficiently solve the optimisation problem in Eq. (3). Note, that the second-order terms in the asymptotic
expansion of stresses and energy, respectively, must be taken into account in the present case to yield an
adequate approximation and accurate failure predictions.
3 Results and discussion
The failure load Ff of the scarf joint specimen under four-point bending loading is depicted in Fig 2(a)
with respect to the interface inclination defined by the angle ψ. For an interface inclination exceeding 85◦ ,
the initiating cracks emanate from the bi-material notch. In this region, failure loads depend significantly
on the interface orientation and the joint strength decreases with increasing values of ψ. The asymptotic
approach yields accurate predictions compared to the experimental results with a maximum relative
error of 9%. Moreover, the relative error between asymptotics and a purely numerical reference solution
is smaller than 6.6%. According to the predictions, cracks tend to initiate perpendicularly to the joints
edge which is reflected by the experiments, cf. 2(b) and 2(c). Reducing the interface angle below 85◦
leads to a transition from notch induced failure to a bending failure mode. Here, cracks initiate vertically
in a certain distance away from the notch tip and failure loads are only hardly sensitive to ψ. To tackle
the problem of bending failure using the coupled criterion, the stress distribution is modelled according
to the elementary Euler–Bernoulli beam theory and the corresponding incremental energy release rate
can be calculated by means of an empirical relation given in [19].
361
4 Conclusion
In this work, failure of a bi-material scarf joint under four-point bending has been investigated by means
of a coupled stress and energy criterion. Depending on the material interface orientation, two distinct
failure mechanisms have been observed, that is, notch induced failure and bending failure, respectively.
Bending failure is characterised by crack onset within the Rohacell foam in a certain distance from the bi-
material notch and failure loads are independent of the interface orientation. In contrast, for sufficiently
large interface angles, crack nucleation is induced by the singular stress field at the notch tip and the
effective joint’s strength depends significantly on the interface orientation. For notch induced failure, the
use of asymptotics enabled an efficient solution of the optimization problem and has led to a significant
reduction of computational effort compared to a fully numerical treatment. Taking higher order terms
into account, accurate predictions of experimentally observed failure loads have been achieved with a
relative error smaller than 7%.
References
[1] M. Burman. PhD thesis. KTH, Stockholm, 1998.
[2] M. L. Dunn, W. Suwito, and S. Cunningham. International Journal of Solids and Structures 34.29
(1997), pp. 3873–3883.
[3] J. Felger, N. Stein, and W. Becker. Engineering Fracture Mechanics 182 (2017), pp. 621–634.
[4] J. Felger, N. Stein, and W. Becker. International Journal of Solids and Structures 122 (2017),
pp. 14–24.
[5] J. Felger, P.L. Rosendahl, D. Leguillon, and W. Becker. International Journal of Solids and Struc-
tures 164 (2019), pp. 191–201.
[6] J. Felger, N. Stein, C. Frey, and W. Becker. Composite Structures 208 (2019), pp. 546–556.
[7] Z. Hashin. Journal of the Mechanics and Physics of Solids 44.7 (1996), pp. 1129–1145.
[8] D Leguillon. European Journal of Mechanics-A/Solids 12.1 (1993), pp. 33–51.
[9] D. Leguillon. European Journal of Mechanics-A/Solids 21.1 (2002), pp. 61–72.
[10] D. Leguillon and E. Sanchez-Palencia. John Wiley & Sons, Inc., 1987.
[11] D. Leguillon and Z. Yosibash. International Journal of Fracture 122.1-2 (2003), pp. 1–21.
[12] Z. Qian and A.R. Akisanya. Acta Materialia 46.14 (1998), pp. 4895–4904.
[13] E.D. Reedy and T.R. Guess. International Journal of Solids and Structures 30.21 (1993), pp. 2929–
2936.
[14] S. Ribeiro-Ayeh and S. Hallström. Engineering Fracture Mechanics 70.12 (2003), pp. 1491–1507.
[15] P. L. Rosendahl and P. Weißgraeber. The Cryosphere Discussions (2019).
[16] P. L. Rosendahl, P. Weißgraeber, N. Stein, and W. Becker. International Journal of Solids and
Structures 113 (2017), pp. 10–23.
[17] G.B. Sinclair, M. Okajima, and J.H. Griffin. International Journal for Numerical Methods in En-
gineering 20.6 (1984), pp. 999–1008.
[18] N. Stein, P. Weißgraeber, and W. Becker. Composite Structures 133 (2015), pp. 707–718.
[19] H. Tada, P. Paris, and G. Irwin. New York: ASME Press 2 (2000).
[20] M.E. Waddoups, J.R. Eisenmann, and B.E. Kaminski. Journal of Composite Materials 5.4 (1971),
pp. 446–454.
[21] P. Weißgraeber, D. Leguillon, and W. Becker. Archive of Applied Mechanics 86.1 (2016), pp. 375–
401.
[22] P. Weißgraeber, J. Felger, D. Geipel, and W. Becker. European Journal of Mechanics-A/Solids 55
(2016), pp. 192–198.
[23] J.M. Whitney and R.J. Nuismer. Journal of Composite Materials 8.3 (1974), pp. 253–265.
[24] Z. Yosibash, E. Priel, and D. Leguillon. International Journal of Fracture 141.1-2 (2006), pp. 291–
312.
362

Hyperelastic adhesive joints:

from crack nucleation to ultimate failure
1
Philipp L. Rosendahl ∗
, Yves Staudt2 , Alexandra P. Schneider1 , Jens Schneider3 and Wilfried Becker1
1 Technische Universität Darmstadt, Fachgebiet Strukturmechanik, Franziska-Braun-Str. 7, 64287 Darmstadt
2 IngenieurbüroBräuer und Späh, Ludwig-Beck-Str. 8, 68163 Mannheim
3 Technische Universität Darmstadt, Institute of Structural Mechanics and Design, Franziska-Braun-Str. 3, 64287 Darmstadt
Abstract. The present analysis proposes a mixed-mode failure model for crack initiation in nonlinear elas-
tic materials. Characterized in independent experiments, the model is used to determine critical loads of
hyperelastic adhesive bonds in both shear and tension dominated configurations. For any of the exam-
ined adhesive joint configurations the model predicts and explains size effects and agrees well with exper-
imental findings. The present work demonstrates how a physical understanding of the failure processes
in complex materials allows for predicting crack nucleation and the derivation of simple engineering tools.
Introduction
The unique molecular structure of silicones provides mechanical properties which render them excellent
structural adhesives. However, at bi-material corners between adherends and adhesive stress singularities
owing to geometrical and material discontinuities are present. In order to capture crack onset at these
singularities, classical approaches such as stress-based criteria or fracture mechanics can only be applied
using an additional length parameter.
In fact, many modern theories of fracture involve a material length scale, i.e., a parameter with the
dimensions of a length. Among these nonlocal approaches there are stress-based methods such as point
method (PM) and line method (LM) where stresses are evaluated either in a point-wise manner in a certain
distance from a stress concentration [14] or averaged over a specific length [13]. Nonlocal approaches using
fracture mechanics are the imaginary crack method (ICM) which examines the energy release rate at the
tip of an assumed pre-existing crack [24] and the finite crack extension method (FCE) which averages the
energy release rate over a finite crack extension [6]. Strain energy density (SED) methods either evaluate
SED along rays from the crack tip [20] or averaged in a finite sized domain [8].
Different nonlocal approaches were successfully applied to hyperelastic materials. Clift et al. [3] evaluate
stress criteria in relatively coarsely discretized finite element analyses (FEAs) of silicone bonds. When a
fixed discretization is used in FEAs, the element size appears as a characteristic length scale implicitly.
Ayatollahi at al. [1] are more specific in their use of a critical distance in rubbers. They evaluate an
effective stretch criterion in a certain distance from notches and obtain good agreements with experiments.
Berto [2] averages strain energy density (SED) in a finite volume around V-notches to predict crack
initiation. Miehe and Schänzel [12] predict crack initiation in hyperelastic materials using the regularized
phase-field approach. The models are able to describe experimental findings on hyperelastic materials
adequately. Yet, the length parameter involved in all of the above approaches is not known a priori and
lacks definite physical meaning. Pipes et al. [15], for instance, find a dependence of the critical distance
on geometrical features which, hence, cannot be regarded a material constant.
Assuming the sudden nucleation of a finite sized crack introduces a length scale with clear physical mean-
ing – the finite size of the initiated crack. In order to determine the finite crack size Leguillon [9] proposed
requiring the simultaneous satisfaction of both a stress and an energy criterion as necessary and sufficient
∗ Corresponding author: rosendahl@fsm.tu-darmstadt.de ORCID iD: 0000-0002-6587-875X
363
a b c ∅140 d
115 steel
8 ∅12
33 2
25 R25 t 12
6 16
R14 steel steel steel
∅170
Dimensions in mm l 12
Figure 1. DOWSILTM 993 specimens used in the present work. a) Dumbbell specimen for uniaxial tests according
to ASTM Standard D412-16, b) H-shaped specimen with adhesive thickness t and bond length l for
simple shear and normal plate separation tests according to ETAG 002, c) tubular lap joint specimen
for unnotched circular shear tests and d) cylindrical specimen for uniaxial compression tests according
to ISO 7743. Please refer to Staudt at al. [21] for details on specimen manufacturing.
condition for crack nucleation. This so-called coupled stress and energy criterion provides excellent pre-
dictions for mixed-mode failure [5, 23], adhesive joint failure [7, 22], the formation of crack patterns [16],
the failure of structures with elliptical open-holes [4, 25] and it can be linked to micromechanical material
properties [11]. A comprehensive review is given by Weißgraeber et al. [26].
Common to all of the above finite fracture mechanics (FFM) studies are the assumptions of brittleness
and linear-elastic material behavior. However, the physical background of the coupled criterion allows
for a generalization of the method. For instance, Leguillon and Yosibash [10] revisit the assumption
of brittleness and develop a theoretical FFM framework for quasi-brittle materials. The present work
elaborates the question of material behavior. It proposes a formulation of finite fracture mechanics
for hyperelastic materials. It uses the example of structural silicone adhesives. Yet, the presented
methodology is applicable to crack initiation from stress concentrations in any nonlinear elastic material.
1 Failure model for brittle crack nucleation in hyperelastic materials
Consider DOWSILTM (a nearly incompressible hyperelastic two-component room temperature curing

silicone) as an exemplary material. Despite large deformations prior to failure, no significant inelastic
deformations are present and DOWSILTM 993 can be considered brittle. Fracture surfaces of a fully
fractured specimen align perfectly because only the formation of the fracture surface dissipates energy.
Using the specimens shown in Figure 1, Staudt et al. [21] characterize the constitutive behavior of
DOWSILTM 993 and Rosendahl et al. [17] develop a strain-based mixed-mode failure criterion for the hy-
perelastic bulk material. Independent fracture toughness measurements by Rosendahl et al. [18] together
with this bulk material failure criterion allow for the derivation of a coupled strain and energy failure
model for corner-induced crack nucleation in notched specimens as proposed by Rosendahl et al. [19].
The criterion assumes the instantaneous formation of a finite sized crack at a critical load. It requires
the simultaneous satisfaction of two necessary conditions: The structure must be overloaded in a finite
domain and the energy release of a finite crack must suffice. The energy condition requires the incremental
energy release rate G to exceed the material’s fracture toughness
∆Π
G(∆A) = − ≥ Gc , (1)
∆A
where ∆Π is the change in total potential energy of the structure with a crack increment of size ∆A.
The nonlinear elastic adhesive is assumed overloaded when an equivalent strain averaged over the entire
potential crack surface εeq (∆A) exceeds a critical threshold
Z
1
εeq (∆A) = εeq (x)dA ≥ εc . (2)
∆A ∆A
The equivalent strain function εeq must be capable of describing bulk material failure under complex
loads at any given location x. Using the second and third invariants IIε0 and IIIε0 of the deviator of the
364
spatial Hencky strain tensor ε† , we obtain an equivalent strain function with the required properties:
" √ !#
p π 1 3 3 IIIε0 (x)
0
εeq (x) = 2IIε (x) cos β − arccos α , (3)
6 3 2 IIε0 (x)3/2
where α and β are shape parameters. For plane problems with through-thickness cracks and out-of-plane
width w we obtain ∆A = w∆a and the coupled criterion reads
εeq (∆a) ≥ εc ∧ G(∆a) ≥ Gc . (4)
Rosendahl et al. [17] report α = 0.86, β = 1.12 and εc = 1.47 ± 0.10 for DOWSILTM 993. Fracture
toughness measurements by Rosendahl et al. [18] yield Gc = 4.22 ± 0.60 N/mm.
2 Prediction of structural failure
In order to validate the failure model, we tested 77 H-shaped DOWSILTM 993 specimens under simple
shear loading, performed normal plate separation tests of 7 H-shaped specimens and circular shear tests
of 4 tubular lap joints. Exemplary results are shown in Figure 2.
Figures 2a and b compare predictions of the FFM failure model to measured failure loads of simple
shear specimens of out-of-plane width w = 12 mm and lengths l = 50 mm and l = 100 mm, respectively.
Experimental data is indicated as mean and standard deviation. FFM failure loads τf and initial crack
lengths ∆a are evaluated using mean fracture properties. In order to account for the uncertainty of
fracture properties, failure load predictions are given with shaded 95% confidence intervals. Overall a good
agreement between model and experiments is observed despite significant scatter in both experimental
data and fracture properties. While failure loads are slightly overestimated, all trends are captured
correctly. Confidence intervals of the predictions overlap with experimental standard deviations in any
case. The experiments show the adhesive thickness effect, i.e., a reduction of the effective joint strength
with increasing adhesive thickness t. The effect originates from an increasing energy release rate with
increasing adhesive thickness which dominates the crack initiation process. Because the coupled strain
and energy criterion accounts for the energy balance, the effect is reproduced well.
Figures 2c and d show the results of normal plate separation test of H-shaped specimens and circular shear
tests of notched tubular lap joints, respectively. We used H-shaped specimens with adhesive thickness
t = 12 mm, length l = 50 mm and width w = 12 mm and tubular lap joint specimens with two 3 mm
notches, adhesive thickness t = 8 mm, width w = 16 mm and inner diameter 140 mm. For the plate
separation test, FFM predictions are slightly conservative. However, as observed in Figures 2a and b,
they agree with experimental standard deviations within their confidence intervals. The circular shear
case shows similar results as the simple shear setup. While in Figures 2c and d, only one geometry of
each adhesive joint is tested, FFM again predicts the adhesive thickness effect which is expected to occur
in the present test configurations as well.
References
[1] M. R. Ayatollahi, M. Heydari-Meybodi, M. Dehghany, and F. Berto. Adv. Eng. Mater. 18.8 (2016), pp. 1364–1370.
[2] F. Berto. Theor. Appl. Fract. Mech. 76 (2015), pp. 17–26.
[3] C. D. Clift, L. D. Carbary, P. Hutley, and J. H. Kimberlain. J. Facade Des. Eng. 2.3–4 (2014), pp. 137–161.
[4] J. Felger, N. Stein, and W. Becker. Eng. Fract. Mech. 182 (2017), pp. 621–634.
[5] J. Felger, P. L. Rosendahl, D. Leguillon, and W. Becker. Int. J. Solids Struct. 164 (2019), pp. 191–201.
[6] Z. Hashin. J. Mech. Phys. Solids 44.7 (1996), pp. 1129–1145.
[7] S. Hell, P. Weißgraeber, J. Felger, and W. Becker. Eng. Fract. Mech. 117 (2014), pp. 112–126.
[8] P. Lazzarin and R. Zambardi. Int. J. Fract. 112.3 (2001), pp. 275–298.
[9] D. Leguillon. Eur. J. Mech. – A/Solids 21.1 (2002), pp. 61–72.
[10] D. Leguillon and Z. Yosibash. Int. J. Solids Struct. 122-123 (2017), pp. 1–13.
√
† The spatial logarithmic Hencky strain tensor reads ε = ln FF> where F is the deformation gradient. Its deviator
tensor is obtained from dev ε = (tr ε)I/3 where I is the identity matrix. In termsof principal strains (εI , εII , εIII ), the second
and third deviatoric invariants read IIε0 = (εI − εII )2 + (εII − εIII )2 + (εIII − εI )2 /6 and IIIε0 = (εI + εII − 2εIII )(2εI − εII −
εIII )(εI − 2εII + εIII )/27, respectively.
365
1.6 8 1.6 8
a
l = 50 mm b l = 100 mm
Nominal shear strength τ f [MPa] Ð→
Nominal shear strength τ f [MPa] Ð→

t t
1.2 ∆a 6 1.2 ∆a 6
Crack length ∆a [mm] Ð→

0.8 4 0.8 4
τf τf
FFM FFM
Experiment Experiment
0.4 ∆a 2 0.4 ∆a 2
0.0 0 0.0 0
0 8 16 24 32 0 8 16 24 32
Adhesive thickness t [mm] Ð→ Adhesive thickness t [mm] Ð→
2.0 10 3.0 12
c d
r i = 70 mm t
∆a
Nominal normal strength σf [MPa] Ð→
∆a t 2.5 10
1.6 8 Nominal shear strength τ f [MPa] Ð→ 3 mm

l = 50 mm
2.0 8
1.2 Experiment 6
FFM σf 1.5 τf 6
0.8 4 FFM
Experiment
1.0 4
0.4 2 ∆a
∆a 0.5 2
0.0 0 0.0 0
9.0 10.5 12.0 13.5 15.0 4.0 6.0 8.0 10.0 12.0 14.0
Adhesive thickness t [mm] Ð→ Adhesive thickness t [mm] Ð→
Figure 2. Finite fracture mechanics (FFM) failure load predictions in comparison to experiments. a) Simple shear
test with overlap length l = 50 mm, b) simple shear test with l = 100 mm, c) normal plate separation test
and d) notched tubular lap shear test. Experimental data is indicated as mean and standard deviation.
FFM failure loads (σf , τf ) are computed using mean fracture properties and are given with shaded 95%
confidence intervals accounting for uncertainty of the fracture properties. Reported finite crack lengths
∆a correspond to mean fracture property predictions.
[11] D. Leguillon, E. Martin, O. Ševeček, and R. Bermejo. Int. J. Fract. 212.1 (2018), pp. 89–103.
[12] C. Miehe and L.-M. Schänzel. J. Mech. Phys. Solids 65 (2014), pp. 93–113.
[13] H. Neuber. Forsch. auf dem Gebiete des Ingenieurwesens 7.6 (1936), pp. 271–274.
[14] R. E. Peterson. In: Stephen Timoshenko Anniv. Vol. New York: Macmillan, 1938, pp. 179–183.
[15] R. B. Pipes, R. C. Wetherhold, and J. W. Gillespie. J. Compos. Mater. 13.2 (1979), pp. 148–160.
[16] P. L. Rosendahl, P. Weißgraeber, N. Stein, and W. Becker. Int. J. Solids Struct. 113-114 (2017), pp. 10–23.
[17] P. L. Rosendahl, M. Drass, J. Felger, J. Schneider, and W. Becker. Int. J. Solids Struct. 166 (2019), pp. 32–46.
[18] P. L. Rosendahl, Y. Staudt, C. Odenbreit, J. Schneider, and W. Becker. Int. J. Adhes. Adhes. 91 (2019), pp. 64–71.
[19] P. L. Rosendahl, Y. Staudt, A. P. Schneider, J. Schneider, and W. Becker. Mater. Des. Under Review (2019).
[20] G. C. Sih. Int. J. Fract. 10.3 (1974), pp. 305–321.
[21] Y. Staudt, C. Odenbreit, and J. Schneider. Int. J. Adhes. Adhes. 82 (2018), pp. 126–138.
[22] N. Stein, P. Weißgraeber, and W. Becker. Compos. Struct. 133 (2015), pp. 707–718.
[23] A. Talmon l’Armée, S. Hell, P. L. Rosendahl, J. Felger, and W. Becker. Eng. Fract. Mech. 186 (2017), pp. 283–299.
[24] M. E. Waddoups, J. R. Eisenmann, and B. E. Kaminski. J. Compos. Mater. 5.4 (1971), pp. 446–454.
[25] P. Weißgraeber, J. Felger, D. Geipel, and W. Becker. Eur. J. Mech. - A/Solids 55 (2015), pp. 192–198.
[26] P. Weißgraeber, D. Leguillon, and W. Becker. Arch. Appl. Mech. 86.1-2 (2016), pp. 375–401.
366

J-integral and mixed-mode cohesive zones

1
Johannes Scheel ⇤
1 University of Kassel, Institute of Mechanics, D-34125 Kassel, Mönchebergstraße 7, Germany
Abstract. The J-integral is a powerful and accurate means for characterizing crack tip loadings and was
independently introduced by Rice [4] and Cherepanov [2]. The calculation of both matrix and interface crack
tip loadings is possible, commonly implementing line or equivalent domain integrals. The crack tip opening
displacement, as fracture mechanics loading quantity of cohesive zone models, on the other hand, was related
to the J-integral by Rice [4], however just for the pure mode I loading case. In this work, mixed-mode loading
conditions are considered, resulting in a generalization of the single mode I relation. The J1 coordinate of the
generalized relation is verified analytically, considering a Dugdale-Barenblatt crack model under mixed-mode
loading.
Introduction
The crack tip loading quantities of linear elastic fracture mechanics are all related to one another. The
J-integral, however, was originally only related to the crack tip opening displacement for the single mode
I loading scenario by Rice [4]. Scheel and Ricoeur [5] showed in their analysis of a matrix crack in
the vicinity of a cohesive interface crack, that in the J-integral vector formulation of the matrix crack,
the cohesive crack tip loading in terms of a J-integral vector can be derived exhibiting two non-zero
coordinates, being the mixed-mode generalization of Rice’s original relation [4]. The generalization of
the J1 coordinate is straightforward and uniquely related to the normal and tangential crack tip opening
displacements, whereas the second coordinate depends on the solution of the boundary value problem
and requires an integration along the cohesive crack faces. The generalized relation is verified using the
Griffith crack model. For the Griffith crack, the J-integral is easily calculated with the stress intensity
factors for single and mixed mode loading and is the reference value. Introducing a cohesive zone in
front of the crack tip, in a Dugdale-Barenblatt crack model [1, 3], the cohesive crack tip loading can
also be calculated for mixed mode loading scenarios using the derived generalized relation. For a ratio
of applied stress to yield stress tending to zero, the length of the cohesive zone also approaches zero.
The small scale yielding approximation is then becoming accurate, so that the J-integral calculated for
the Dugdale-Barenblatt crack model with the generalized relation approaches the reference value. This
verification is restricted to the J1 coordinate in this work.
1 The J-integral vector of a mixed-mode loaded cohesive zone
The generalized relation for a mixed mode loaded cohesive crack given by Scheel and Ricoeur [5], shows
the relation between J-integral and crack tip opening displacements in the first coordinate of the vector
and the dependence on the boundary value problem in the second coordinate:
0 t t
1
Z2 Z1
B C
Jk = @ t2 ( 2 )d 2 + t1 ( 1 )d 1 A e~1
0 0
0 1 (1)
Z l
1
+@ ( + +
11 "11 11 "11 + 22 ("22 "+
22 ) + 2 12 (⌦12 ⌦+
12 ))dx1
A e~2 .
2
0
⇤ Corresponding author: j.scheel@uni-kassel.de ORCID iD: https://orcid.org/0000-0003-4400-0283
367
Here, δit = δ1/2

t
describe the opening separations at the physical crack tip and it is assumed that the
tractions ti are only functions of the mode-associated separations δi . The integration contour has been
shrunk to the cohesive surfaces of a straight cohesive crack [4], with a cohesive zone length of ∆l and
the superscripts +/− indicate affiliations of the quantities to the upper (+) or lower (−) fictitious crack
+/−
face. The tangential strain ε11 is not continuous and thus has to remain separated, while the tractions
− +/−
are continuous, i.e. t+
i = −ti . This formulation involves the jumps of normal strain ε22 and rotation
+/−
Ω12 over the cohesive crack faces, with
1
Ωij = (ui,j − uj,i ), (2)
2
where ui,j is the displacement gradient.
2 Verification of the first coordinate J1
For the verification of the first coordinate of the generalized relation of Eq. (1), it is assumed that the
cohesive laws for mode I and II are equal, that the constitutive behavior is equal to perfect plasticity and
that the cohesive zone is equally mode I and mode II loaded. For the calculation of the first coordinate of
the J-integral vector of Eq. (1) the crack tip opening displacements are needed, but in order to calculate
them the length of the cohesive zone needs to be calculated first. With the assumptions made, it is
calculated exploiting the condition that there is no stress singularity at the fictitious crack tip, yielding
∞
πσ
∆l = a sec −1 , (3)
2σ0
where a is the Griffith crack length, σ ∞ is the applied stress and σ0 is the yield stress. The crack tip
opening displacements then can be calculated as

t σ0 a a + ∆l
δi = 8 ln , (4)
πE 0 a
where E 0 = E is the Young’s modulus for plane stress and for plane strain E 0 = E/(1 − ν 2 ), with ν as
Poisson’s ratio ans E as Young’s modulus. The J1 coordinate for the Dugdale-Barenblatt crack model,
in the following denoted as J1coh , is calculated inserting Eqs. (4) and (3) into the first coordinate of Eq.
(1), yielding
∞
πσ
J1coh = 16 σ02 ln sec . (5)
2σ0
The J1 coordinate of the Griffith crack, in the following denoted as J1LEF M , can be calculated with the
stress intensity factors by the established relation,
KI2 + KII
2
2(σ ∞ )2 πa
J1LEF M = 0
= . (6)
E E0
The ratio of the two calculated values is plotted in Figure 1 versus the ratio of applied stress to yield
stress. The ratio of the Dugdale-Barenblatt crack length (a+∆l) to the Griffith crack length a is depicted
there too. It is clearly visible that for a stress ratio going to zero, the calculated coordinates J1coh and
J1LEF M as well as the Griffith and the Dugdale-Barenblatt crack lengths become equal. A very small
cohesive zone size complies with the small scale yielding approximation, therefore J1coh being equal to
J1LEF M for that condition, verifies the first coordinate J1coh of the generalized relation of Eq. (1).
3 Conclusion
The J-integral vector of a mixed mode loaded cohesive zone derived by Scheel and Ricoeur [5], being
the generalization of the scalar formulation derived by Rice [4], is picked up in this work. The second
coordinate of the vector is not as straightforwardly derived as the first. In contrast to the first coordinate,
the second requires the solution of the boundary value problem, whereas the first coordinate has a direct
relation to the cohesive laws and the crack tip opening displacements. The first coordinate is verified in
368
1.30
[−]
1.25 a + ∆l
a
1.20 J1coh
J1LEF M
1.15
1.10
1.05
1.00
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

σ ∞/σ0 [−]
Figure 1. Ratio of the calculated J1 coordinates and ratio of the Dugdale-Barenblatt crack length to the Griffith
crack length, plotted vs. the ratio of applied stress to yield stress
this work at a Griffith crack model. The J-integral classically calculated with the stress intensity factors
for the Griffith crack is accurate with the condition of small scale yielding and acts as reference. For a
Dugdale-Barenblatt crack with a cohesive zone length tending to zero, this condition is also valid and in
the limiting case the model becomes equal to the Griffith crack model. It is shown, that the J1 coordinate
calculated for the Dugdale-Barenblatt crack with the generalized relation, equals the J1 coordinate of the
Griffith crack for the small scale yielding approximation, finally confirming the generalized relation for
the first coordinate. The verification of the second coordinate is in progress.
References
[1] G. I. Barenblatt. The mathematical theory of equilibrium cracks in brittle fracture. Advances in
applied mechanics 7 (1962), pp. 55–129.
[2] G. P. Cherepanov. Crack propagation in continuous media (translation from russian). Journal of
Applied Mathematics and Mechanics 31 (1967), pp. 503–512.
[3] D. S. Dugdale. Yielding of steel sheets containing slits. Journal of the Mechanics and Physics of
Solids 8 (1960), pp. 100–104.
[4] J. R. Rice. A path independent integral and the approximate analysis of strain concentration by
notches and cracks. Journal of Applied Mechanics 35 (1968), pp. 379–386.
[5] J. Scheel and A. Ricoeur. A comprehensive interpretation of the J-integral for cohesive interface
cracks and interactions with matrix cracks. Theoretical and Applied Fracture Mechanics 100 (2019),
pp. 281–288.
369
370

Finite element model to describe crack deflection and

constrained micro-cracking
1
Markus Tauscher ∗
, Martin Pletz1 , Florain Arbeiter2 and Clara Schuecker1
1 Designing Plastics and Composite Materials, Department of Polymer Engineering and Science, Montanuniversitaet Leoben,
Austria
2 Materials Science and Testing of Polymers, Department of Polymer Engineering and Science, Montanuniversitaet Leoben, Austria
Abstract. An incremental finite element model is presented that predicts the growth of a crack through a 2D
structure containing round pores. The crack is deflected due to the influence of the pores on the strain field
and might grow into a pore. Furthermore, the initiation of cracks from the pore surfaces is modelled with a
combined stress-energy criterion. The model contributes to the understanding of toughening mechanisms in
heterogeneous materials. It is capable of showing the influence of the pore content, the pore arrangement and
pore shape on the resistance to crack propagation.
1 Introduction
In some heterogeneous materials, exceptionally high strength and toughness values are observed. It is well
known that the strength and toughness is not only determined by the properties of the constituents, but
also by the arrangement of the inclusions in the matrix. For an existing crack, effects like fiber bridging,
crack deflection, residual stresses and constrained micro-cracking can reduce the crack’s driving forces
and thus increase the composites toughness. Through this crack-arresting effect, hierarchical materials
can also be less sensitive to existing cracks, making them more damage tolerant. Hierarchical materials,
as found in nature, can feature such toughening mechanisms on different length scales [5]. Typical exam-
ples for such hierarchical structures are nacre, bone, and wood. Using those mechanisms for engineering
materials seems very promising, but is often limited by the capabilities of the manufacturing processes.
In recent years, however, production technologies have been rapidly evolving and make it possible to
manufacture more complex structured compounds at reasonable costs.
Many groups of engineers work on understanding those toughening mechanisms in natural materials [3]
and using such mechanisms to develop engineering materials with improved strength and toughness [4]
by means of testing, modelling and manufacturing.
In this work, a numerical model to predict crack deflection and constrained microcracking in inhomo-
geneous materials is developed. By doing so, both crack propagation and initiation must be accounted
for. While crack propagation in brittle bulk materials is well established (Griffith–Irvine), some novel
concepts such as the Finite Fracture Mechanics (FFM) concept [2, 4] can predict crack initiation. This
concept is based on coupling a stress and an energy criterion, which both have to be fulfilled for a crack
to initiate. Here, a Finite Element (FE) model for a Compact Tension (CT) specimen with an initial
crack and several inclusions is used. These inclusions can deflect, trap or initiate the crack. As a first
step towards describing some of these toughening mechanisms, circular voids are considered rather than
inclusions to avoid competing fracture mechanisms like delamination and other effects of the interface.
Such specimen can easily be manufactured through 3D printing, which is currently under way.
2 Finite element model setup
The FE model represents a CT specimen that is used for testing fracture toughness and is set up in
the finite element code ABAQUS [1]. Figure 1a shows the model’s geometry and the applied boundary
∗ Corresponding author: markus.tauscher@unileoben.ac.at
371
conditions. The surfaces of the two left holes are kinematically coupled to reference points, where a
displacement load u is applied. Those reference points are constrained in the horizontal direction but
free to rotate.
The model is set up with linear geometry and uses quadratic plane stress elements with reduced inte-
gration. The material used in the numerical analysis is polylactic acid (PLA), which is quite common in
3D printing. The selected PLA features
√ a Young’s modulus of 3.4 GPa, a Poisson’s ratio of 0.37 and a
fracture toughness KIc of 4.5 MPa m. Typical strength values of PLA are about 40 MPa. For plane
2
stress, the fracture energy Gc is obtained as Kc2 /E = 5.96 mJ/mm . The analysis assumes the material
to behave in a brittle manner, which is valid for PLA if the temperatures are low and the loading rate is
high. Between the crack faces, no contact is defined because in the first simple configuration, no closed
and compressed cracks occur.
At the beginning of the analysis, there is an initial crack with a length of 10 mm in the center of the spec-
imen. This is the primary crack. During loading, additional secondary cracks can initiate from the pores.
The displacement load is applied in steps of ∆u of 0.2 mm. In each increment of the modelling scheme,
see Figure 1b, the FE model containing all existing cracks is run and crack initiation from the pores is
checked. In the next step, the primary crack and all initiated cracks are checked for crack propagation.
If crack growth is predicted for those cracks, the cracks are extended and the loop starts again. Once no
cracks initiate and no cracks grow in one increment, the displacement load is increased by ∆u and crack
initiation and crack growth is checked for this increased load similar as for the previous displacement.
u/2 initiated crack

25 FE model with
crack list crack initiation
r=3 no
ui+1 = ui + Du Pores: max. s reached?
2
22
50
initial crack yes

no crack propagation no FE model with crack 
Cracks: G > Gc? to surface: G > Gc?
10 30
20
u/2 yes
yes
50 propagate crack by Da
initiate new crack
(a) (b)
Figure 1. Geometry (dimensions in mm) and loads applied on the CT specimen model (a) and work flow of the
incremental modelling method (b).
2.1 Crack initiation from pores
The crack increment ∆a is taken as a suitable value that can yield a sufficiently fine discretization of
the crack growth. For crack initiation, the maximum principal stress is evaluated at a normal distance
∆a from the pore surface, and the position of highest stress along the circumference of the pore is taken
as the critical location. If this stress is higher than the material’s strength, a crack of the length ∆a is
initiated at that location normal to the surface. Then, it is checked whether the J-integral of this crack
is above the material’s fracture energy. This is a simplified formulation of the combined stress/energy
criterion for crack initiation, where both energy release rate and stress are evaluated at different positions
starting from the surface, and the crack initiation with different crack lengths can be predicted. It should
be noted that the strength used in the FFM stress criterion is not the actual strength of the material but
has to be calibrated. In this study, it was assumed to be 100 MPa, but will be calibrated when first test
results are available.
2.2 Crack growth
In each iteration, the J-integral of the cracks is evaluated and used for assessing the crack’s growth by
comparing it to the fracture energy Gc . The maximum tangential stress criterion as implemented in
ABAQUS [1] is used to obtain the direction of crack propagation, as done in a similar analysis in [6].
372
When a crack tip comes close to a pore surface such that the normal distance is less than the value of
∆a, the crack is extended into the pore. In the future, more sophisticated concepts for predicting crack
growth such as the configurational force concept will be studied.
3 First model results
The modelling scheme is used for a CT specimen with and without a pore. The crack length increment
∆a is chosen as 0.04 mm. For the model that contains a pore, there is a vertical offset of the pore from the
crack plane, as indicated in Figure 1a. Figure 2 shows contour plots of the maximum principal stresses in
the CT specimen as the crack grows. The instants of the contour plots are taken for a horizontal length
of the crack ax of 17.6 mm, 22.3 mm and 31.6 mm in Figure 2a, b and c, respectively. There is a distinct
deflection of the crack from its initial plane as it grows towards the pore. When the crack grows into the
pore, it is arrested and has to initiate again. Note the reduced level of maximum stress when the crack
has grown into the pore in Figure 2b.
(a) (b) (c) (d)
Figure 2. Contour plots of the maximum principal stresses in the model with a crack length of a) 17.6 mm, b)
22.3 mm and 31.6 mm. The legend of the contour plots (in MPa) is given in (d).
To assess the influence of the pore on the crack’s propagation, Figure 3 plots the horizontal crack length
ax over the applied displacement load u. For the CT specimen without a pore, the initial crack starts to
grow at a value of u=0.6 mm and leads to a gradual failure of the specimen. The specimen with pore
has an earlier growth of the initial crack (at a displacement u of 1 mm, the crack already runs into the
pore), but then stops and has to re-initiate at the pore, which it does at a value of u=2.2 mm. Apart
from the region with a displacement load u of 1 mm to 2.2 mm, there are only slight differences between
the two curves.
Figure 3. Comparison of model results between a CT specimen without and with pore in terms of
load/displacement u vs. horizontal crack length ax . There is an influence of the pore in the region
of the pore: Towards the pore, the crack grows earlier, and then, a higher u value than in the model
without pore is needed for a further crack extension.
373
4 Conclusions
The shown modelling procedure is capable to predict the influence of a pore on the crack propagation
in a CT specimen. The crack is deflected by the pore and has to re-initiate after growing into the pore.
For the selected geometry and material parameters, the overall course of the crack length – displacement
curve is similar for the models with and without pore. In the future, the modelling approach can be
extended towards more sophisticated methods for crack growth and crack initiation.
For the validation, CT specimens will be 3D-printed and tested. During testing, strains and crack
paths are captured by Digital Image Correlation and can be compared directly to the results of the
model for each load step. The validated model can help to understand crack deflection and constrained
microcracking to identify key parameters and design rules. Furthermore, the model can be extended to
multiphase compounds, where interface failure will play a role.
References
[1] Abaqus. ABAQUS Theory Guide, 6.14. Providence, RI, USA: Dassault Systèmes, 2014.
[2] Pietro Cornetti, Nicola Pugno, Alberto Carpinteri, and David Taylor. Finite fracture mechanics:
A coupled stress and energy failure criterion. Engineering Fracture Mechanics 73.14 (Sept. 2006),
pp. 2021–2033. issn: 00137944. doi: 10.1016/j.engfracmech.2006.03.010.
[3] Zian Jia and Lifeng Wang. 3D printing of biomimetic composites with improved fracture toughness.
Acta Materialia 173 (July 2019), pp. 61–73. issn: 13596454. doi: 10.1016/j.actamat.2019.04.052.
[4] D. Leguillon, E. Martin, O. Ševeček, and R. Bermejo. Application of the coupled stress-energy
criterion to predict the fracture behaviour of layered ceramics designed with internal compressive
stresses. European Journal of Mechanics - A/Solids 54 (Nov. 2015), pp. 94–104. issn: 09977538. doi:
10.1016/j.euromechsol.2015.06.008.
[5] Flavia Libonati and Markus J. Buehler. Advanced Structural Materials by Bioinspiration: Advanced
Structural Materials by Bioinspiration. Advanced Engineering Materials 19.5 (May 2017), p. 1600787.
issn: 14381656. doi: 10.1002/adem.201600787.
[6] Z. Majer, M. Pletz, C. Krautgasser, L. Náhlík, P. Hutař, and R. Bermejo. Numerical Analysis of
Sub-critical Crack Growth in Particulate Ceramic Composites. Procedia Materials Science 3 (2014),
pp. 2071–2076. issn: 22118128. doi: 10.1016/j.mspro.2014.06.335.
374

Stress Control on Interfaces Using Shape Optimisation

Felix Wohlgemuth ∗1
1 Structural Mechanics, TU Dortmund University, Dortmund, Germany
Abstract. The stress field information states a main point of interest, when regarding structural optimisation
of bimaterial structures. The introduction of stress criteria along the volume is crucial in the development of
new designs. In the case of bimaterials the stress field in particular along the interface deserves additional
attention e.g. to prevent delamination. Tracing the interface through shape optimisation in CAD-based
methods is rather expensive due to the high cost in remeshing techniques. Therefore, level set methods with
fixed background meshes are used as in [5].
In this work an approach for controlling strains and stresses on the interface in a framework of shape op-
timisation is introduced. The geometry description is managed by parametrised level set functions and a
sub-meshing technique is coupled with the extended finite element method. The parametrisation with su-
perellipses allows to reduce the number of design variables to a minimum of six variables per introduced
ellipse while holding up a sufficient precision in the geometry description, c.f. [4]. Moreover, it simplifies
the shape derivatives as it provides an implicit description for the moving interfaces. The sub-meshing tech-
nique makes it possible to keep existing strategies from homogeneous structures and to transform them on
a discontinuous material using enriched shape functions provided by the standard extended finite element
method, c.f. [1]. Shape sensitivities are evaluated on the sub-elements and extrapolated to the interface
introducing pseudo nodes. The sensitivity information of the stress field among these pseudo nodes can be
used in the framework of stress minimisation as well as for a side condition in a volume minimisation setup.
Introduction
In many cases the interface between different materials is the weak point of a structure. Stress-induced
phenomena like plasticity or delamination occur in particular there. Thus, an investigation of the stress
field along the interface is of special interest for the engineer. And a main goal in the development of new
designs is to control the stress state at the material discontinuity. This work will focus on the restriction
of stresses with the tools of structural optimisation. The concept is to change the shape and size of the
interface in order to find the optimal stress state. For this a modified extended finite element formulation
is used for both the structural and the sensitivity analysis.
1 A Modified XFEM Approach
Discontinuous material behaviour in the framework of standard finite element methods requires a mesh-
ing of the structure following the material interface. With a moving discontinuity, remeshing in every
optimisation step would be needed. Therefore, an alternative approach is used.
The extended finite element method (XFEM) was developped to represent discontinuities in the field
quantities with a fixed mesh, c.f. [3]. The unsteady behaviour within the element is modelled by enhanc-
ing the shape functions with a part that represents the particular behaviour
X N X
X
uh (x) = Ni (x) ui + Ni∗ (x) ψ j (x) aji . (1)
i∈I j=1 i∈Ij∗
In this work a modified approach is used, which divides the cut element into sub elements alining with
the interface within the element. Those uncut sub elements are treated as standard finite elements. The
field quantities of the sub nodes which emerge during the sub meshing are then extrapolated to the global
∗ Corresponding author: felix.wohlgemuth@tu-dortmund.de
375
nodes. The stresses are evaluated on the gaussian points in the sub elements and weighed onto the sub
nodes using the sub element area. Along the intersection points of the interface with the element edges,
pseudo nodes are introduced. The stresses along the interface can then be computed as
P
σip dAip
σint = P , (2)
dAip
where the integration point stresses (σip ) are calculated with standard finite element techniques within
one sub element and the desired interface stresses (σint ) are merged from these.
2 Structural Optimisation
In this work, two different concepts for the stress control are introduced. On the one hand we use the
stress as the objective function and aim to minimise the interface stresses themselves while holding up
to a prescribed volume of the structure. For this approach we need to measure the stresses along the
interface in order to reduce the cost of the optimisation algorithm. This is done by the p-norm. On the
other hand the stresses are constraint through upper (or lower) bounds while the volume is minimised.
Z 1/p
min V (s), min J(s), with J(s) = p
σ(s) dx(s) ,
s∈D
(3) s∈D ∂Ω (4)
s.t. σ(s) ≤ σmax
s.t. Vincl (s) ≤ Vmax
For both of these concepts the sensitivities of the stresses with respect to changes in the design, i.e. the
shape and size of the interface are required. We keep in mind that the design of the interface is defined
by a superellipse, providing six parameters which are our design variables s. The sensitivity is now given
by the total variation of the integration point stresses with respect to the design
∂uej dū ∂xej
δs σip = δuej σip + δxej σip
∂ ū ds ∂s
(5)
TS
= Au TS + Ax W = [Au , Ax ] .
W
It is noted that the variation is split up to a sub elemental part including the variation of stresses
with respect to sub nodal quantities (Au , Ax ) and a mapping part for the displacements (T S) and the
coordinates (W), see [6] aswell. This seperation makes it possible to use existing techniques for the
sensitivity analysis on sub element level, for example the variational approach from [2], which states the
main benefit of this approach in contrast to other XFEM-based optimisation techniques.
References
[1] Franz-Joseph Barthold and Daniel Materna. A modified extended finite element method approach
for design sensitivity analysis. International Journal for Numerical Methods in Engineering 104.3
(2015). nme.4930, pp. 209–234.
[2] Franz-Joseph Barthold, Nikolai Gerzen, Wojciech Kijanski, and Daniel Materna. Efficient Variational
Design Sensitivity Analysis. In: Mathematical Modeling and Optimization of Complex Structures. Ed.
by Pekka Neittaanmäki, Sergey Repin, and Tero Tuovinen. Cham: Springer International Publishing,
2016, pp. 229–257.
[3] Nicolas Moës, John Dolbow, and Ted Belytschko. A finite element method for crack growth without
remeshing. International Journal for Numerical Methods in Engineering 46.1 (1999), pp. 131–150.
[4] Lise Noël, Laurent Van Miegroet, and Pierre Duysinx. Analytical sensitivity analysis using the
extended finite element method in shape optimization of bimaterial structures. International Journal
for Numerical Methods in Engineering 107.8 (2016). nme.5181, pp. 669–695.
[5] Stanley Osher and James A. Sethian. Fronts propagating with curvature dependent speed: algorithms
based on Hamilton–Jacobi formulations. Journal of Computational Physics (1988), pp. 12–49.
[6] Felix Wohlgemuth and Franz-Joseph Barthold. Structural Optimisation Using Parametrised Level
Set Methods and Extended Finite Element Method. PAMM 18.1 (2018), e201800367.
376
MS 19: Brittle Fracture Analysis in Solids: Experiment and Simulation

Fragment Detection Algorithm for Postprocessing High

Velocity Impact and Natural Fragmentation
1,2
Marvin Becker ∗
, Marina Seidl1 , Miriam Mehl2 and Mhamed Souli3
1 ISL (Institute of Saint Louis), 5 Rue de General Cassagnou, BP 70034, 68300 Saint-Louis - France
2 University of Stuttgart, Universitätsstr. 38, 70569 Stuttgart - Germany
3 Lille 1 University, Cité Scientifique, 59650 Villeneuve d’Ascq - France
Abstract. Post-processing software like Paraview and LS-Prepost are valuable tools for the visualization of
simulation output. However, when fracture occurs it is not possible to distinguish separated parts in the
output. In particular, for evaluating the mass or average velocity of fragments the manual workaround is
time-consuming and error-prone.
There is only little literature available on this particular application, but the problem can be abstracted from
graph theory. If the nodal positions of the particles are known, a suitable criterion determining the graph
connectivity is needed. An algorithm from literature is used to determine the connectivity graph and thus
the fragments. As the number of nodes in a three-dimensional simulation is very large an efficient algorithm
and a fast implementation is indispensable. A further technical challenge is the incorporation of the fragment
information back into the visualization data.
The algorithm is applied to two applications. First, a high-velocity impact of a steel cylinder with round shaped
nose is investigated. Second, the natural fragmentation of a steel cylinder shell under blast is visualized with
the tool. Besides the fragment visualization, a statistical investigation of the fragments shows the versatility
of this code. Experimental results of the two applications have been conducted in the laboratories of ISL
(Institute of Saint-Louis) for validation purposes. Thus, the approach gives the possibility to quickly evaluate
if the numerical prediction and the experimental results coincides based on the statistical information of
the fragments. The temporal crack evolution or formation of fragments can only be investigated using the
numerical results processed with this methodology.
Introduction
Numerical simulations are used in structural analysis to investigate damage and fracture. In some ap-
plications, like durability studies or crash tests, the interest is restricted to the binary information if the
specimen stays intact. However, for other applications the fracture behavior itself needs to be examined
explicitly. For the development of advanced protective structures, a predictive model, which describes
the fracture behavior of a ballistic threat properly, is important. To validate such a numerical model,
an academic example with similar properties is the natural fragmentation of a cylinder shell under blast.
These two examples are investigated with the new fragment detection approach in this abstract. The
field of applications where researchers are interested in the emerging fragments is much broader, as the
following example from aerospace engineering shows: During the takeoff and touchdown of an airplane
the blades of the turbine engine can fracture if the local loads on them are too high. It is important that
these fragments do not exceed a certain size and velocity, such that they do not harm the engine [5].
To include a tracking of fragments in the post processing requires a change of the data structure and is
thus not available in most of the post processors - for instance, not in those used by the authors. To
overcome this shortcoming, two different approaches to detect fragments in particle simulations based on
the nodal positions are presented in the following. This extended abstract summerizes the underlying
theory and algorithmic aspect, and shows an extract of results visualized with the software MegaMol,
developed at the VISUS (University of Stuttgart) [6].
∗ Corresponding author: marvin.becker@isl.eu
377
1 Theory
From a mathematical point of view, fragment identification can be considered as a question in graph
theory. Starting with an arbitrary set of particles (nodes), connections (edges) are created between nodes
which are close enough to each other, resulting in an undirected graph. The fragments are then defined
as the connected components of this graph. This section summarizes the two approaches used to identify
fragments based on a point cloud in three dimensions.
1.1 Problem Description
Let S be a set of points in R3 × R, then two particles xi ∈ S and xj ∈ S are referred to as neighbors at
time t0 if and only if
||xi (t0 ) − xj (t0 )|| < ε, (1)
where || · || denotes the Euclidean norm and ε is the threshold distance. The particles are called connected
neighbors if and only if they are neighbors at all times t:
||xi (t) − xj (t)|| < ε ∀t. (2)
A fragment is the recursive set of connected neighbors. Let Xi be the set of points belonging to a fragment
which contains the start point xi , and cn(xj ) be the function which returns the connected neighbors of an
arbitrary point xj . Then, the function determining all points can be expressed in pseudo-code as follows:
def add_frag(Xi , xi ) :
Xi = {Xi , cn(xi )\Xi }
Xi = add_frag(Xi , xj ) ∀xj ∈ {cn(xi )\Xi }}
return Xi
Xi = add_frag(xi , xi )
1.2 Classical Approach
The classical approach solves the problem described in the previous section straight forward. First, the
neighbors of all particles are determined: This can be done with different algorithms: Comparing all
particle with each other gives a complexity of O(n2 ). One possibility to reduce the best-case complexity
to O(n), is to discretize the particle domain with voxels and check for neighbors only in neighboring
voxels. If the particle distribution is very sparse, such a discretization can be very memory consuming,
or again, result in a worst-case complexity of O(n2 ). Instead, a more advanced discretization of voxels
is needed, which is beyond the scope of this work, and thus, an alternative approach is suggested: If the
neighbor list is generated for the initial time-step, the particles are densely packed, much less memory is
required, and best-case complexity can be ensured. Zhang, even circumvents generating the neighbor list
at all, whilst using initial particle positions based on a FEM mesh where the neighbors are given by the
structured mesh [9]. It has been shown that such a particle distribution is inferior for its uniformity [1].
Thus, this approach is not applied, and instead the neighbor list is generated for the initial time step to
allow a more general particle distribution. The neighbor list, or graph connectivity, is then updated for
the following time-steps and connections which do not exist further are deleted.
Finally, the connected components of the resulting undirected graph are identified using either breadth-
first search or depth-first search, which are both linear in time.
1.3 Machine Learning Approach
Another possibility to determine fragments from a set of points is using unsupervised machine learning.
Density-based spatial clustering of applications with noise (DBSCAN) is a data clustering algorithm
introduced by Ester et al. [4]. The algorithm is calibrated with two parameters: the minpts value defines
the amount of connections required to belong to a cluster of points, and the eps value determines whether
or not two points are connected. A development of this algorithm is HDBSCAN, where the H stands for
hierarchical. This modification determines the fitting parameters automatically: Different values of eps
are used and the result is integrated to give the best stability over eps.
378
(a) t=0 µs. (b) t=40 µs. (c) t=80 µs. (d) t=120 µs.
Figure 1. Visualization of high-velocity impact (800 m/s, 30◦ NATO obliquity). The largest fragment is visualized
in the initial position at four different time steps. In particular, the crack formation can be identified
with this visualization technique (highlighted in red).
2 Results
This section shows the possibilities which emerge when applying one of the proposed techniques to post
process the simulation output. Two different applications are investigated: The first application is an
oblique high-velocity impact of a steel cylinder with round-shaped nose. Here, the aim is to process the
largest fragment, to compare it to experimental data. The second application is the natural fragmentation
of a cylinder ring subjected to blast. For this application, a statistic evaluation of the fragment-size-
distribution is required for the validation of the numerical method.
2.1 High-Velocity Impact
During a high-velocity impact large pressure, strain rate and stresses emerge [7]. Depending on the impact
scenario - impact velocity, obliquity angle, yaw of the impactor, and the involved materials - the material
fails, leading to fracture or partial fragmentation of the impactor[8]. Especially for the case of oblique
impact, the bending results in localized stresses and provokes such mechanisms. From an experimental
point of view, recovering the largest fragment can be realized easily and is therefore well suited for a
comparison. The mass, shape, and velocity of this fragment are properties which can be tested during
the experiment. Having identified the fragments in the numerical simulation provides all these properties
automatically to quickly evaluate parameter studies in a consistent manor.
Figure 1 shows the largest fragment of an oblique impact at different time steps. Besides the properties
which are tested post-impact, the crack evolution can be observed in the time series of the images.
2.2 Natural Fragmentation of Cylinder Shells
The second example where the interest lies in an evaluation of

fragment size, fragment velocity, and a visualization of crack
patterns is the natural fragmentation of explosively driven
cylinder rings [3]. While the fragment-size-distribution can
be tested a posteriori in the experiment, the physical mech-
anisms can only be revealed with numerical simulation: be-
cause of the large deformations of the fragments during the
ring expansion, the ring can not be reconstructed. To argue
that the simulation is reproducing the experimental behav-
ior well, the fragment-size-distribution can be compared as a
measure of accuracy. Since there is no direct post processing
option for this task, the proposed algorithm fills the gap to Figure 2. Evaluation of the fragment veloc-
ity over the initial axial position
get the distribution and visualize the crack patterns (com-
and comparison to the experiment.
pare Figure 2 and 3). The particle visualization software Each scatter point visualizes one
MegaMol [6], is used for the three-dimensional visualization fragment. [2].
of particles.
379
(a) t=0 µs. (b) t=50 µs. (c) t=100 µs. (d) t=150 µs.
Figure 3. Visualization of natural fragmentation with MegaMol. Each color represents a different fragment. Due
to a large amount of fragments similar colors might appear.
3 Outlook
Hitherto, the detection and visualization of the fragments is implemented in a standalone application.
The aim is to exchange this knowledge with the developers of post processors like LS-Prepost, to see this
being developed for a wider audience.
Acknowledgments
The authors acknowledge the contribution of the VISUS department at University Stuttgart for setting
up MegaMol, the Institute of Saint-Louis (ISL) for funding the PhD, and the Universities of Lille and
Stuttgart for supervising the PhD.
References
[1] Marvin Becker, Marina Seidl, Miriam Mehl, and Mhamed Souli. A Study on Initial Conditions in
Smooth Particle Hydrdynamics Modelling of High-Velocity Impact. In: 2018 Nordic LS-DYNA User
Conference. 2018.
[2] Marvin Becker, Tom De Vuyst, Marina Seidl, Miriam Mehl, and Mhamed Souli. Numerical And Ex-
perimental Evaluation of Natural FragmFragment of Explosively Driven Cylinder Rings‘. Procedings
of Light Weight Armor Group 2019 (tbp) (2019).
[3] Tom De Vuyst, Rade Vignjevic, James C Campbell, Andreas Klavzar, and Marina Seidl. A Study of
the effect of aspect ratio on fragmentation of explosively driven cylinders. Procedia Engineering 204
(2017), pp. 194–201.
[4] Martin Ester, Hans-Peter Kriegl, Joerg Sander, and Xiaowei Xu. A density-based algorithm for
discovering clusters in large spatial databases with noise. Prodeedings of the Second International
Conference on Knowledge Discovery and Data Mining (KDD-96) (1996).
[5] Norbert Faderl, G Guetter, Th Wolf, C Doenlen, and S Gaisser. Modellversuche zur Simulation des
Impakts von Turbinensplittern aus der Nickel-Basis-Legierung Inco 718 auf strukturierte Triebwerk-
steile. Tech. rep. French-German Research Institute Saint-Louis (ISL), 2008.
[6] Sebastian Grottel, Michael Krone, Christoph Muller, Guido Reina, and Thomas Ertl. MegaMol -
A Prototyping Framework for Particle-Based Vizualization. IEEE Transactions on Visualization
(2015).
[7] Zvi Rosenberg and Erez Dekel. Terminal ballistics. Springer, 2012.
[8] Marina Seidl, Thomas Wolf, and Rainer Nuesing. Numerical investigations on ricochet of a spin-
stabilised projectile on differently shaped target surfaces. In: 30th ISB, Long Beach. 2017.
[9] Xiaotian Zhang, Jia Guanghui, and Hai Huang. Fragment identification and statistics method of
hypervelocity impact SPH simulation. Chinese Journal of Aeronautics 24.1 (2011), pp. 18–24.
380

Multiphase-eld Simulation Study of Brittle Anisotropic

Crack Propagation in Sandstones
∗ †1 †1 †1 1,2 1,2
Michael Späth , Nishant Prajapati , Christoph Herrmann , Daniel Schneider , Michael Selzer and
1,2
Britta Nestler
1 Institute of Applied Materials (IAM-CMS), Karlsruhe Institute of Technology (KIT), Straÿe am Forum 7, 76131 Karlsruhe,
Germany
2 Institute of Digital Materials Science (IDM), Karlsruhe University of Applied Sciences, Moltkestrasse 30, 76133 Karlsruhe,
Germany
Abstract. This work outlines a computational framework to model anisotropic transgranular and intergranular
fracture propagation in sandstones at microscopic length scale using a multiphase-eld. We incorporate an
anisotropic crack resistance formulation to account for the variations in the fracture toughness of quartz along
dierent crystallographic directions. Moreover, in order to model the intergranular fracturing, a reduced
interfacial crack resistance formulation is utilised. We demonstrate the capabilities of the model through the
representative numerical example.
Introduction
Fractures are ubiquitous in rocks such as sandstone and originate due to a wide range of mechanical
loading conditions. At microscale, quartzarenite sandstone primarily consists of quartz grains with dif-
ferent crystallographic orientations. The intergranular pore space is lled with quartz cements that are
deposited in optical continuity with the grains. Experiments [1] report that the fracture toughness of
prismatic quartz varies with the crystallographic direction, giving rise to preferential direction of crack
growth inside each grain. Furthermore, when the tip of a propagating fracture reaches a grain boundary,
depending upon the resistance oered in that region, intergranular, transgranular or mixed fracturing
might occur. All these anisotropies at grain scale, along with the loading conditions, control the crack
path and account for complex fracture geometries at microscales. These microcracks inuence the phys-
ical properties of rocks, like strength and permeability, and act as a nucleation site for marcofractures.
Therefore, a deep understanding of the formation and growth of microfractures is imperative.
Computational modeling of fracturing in geological materials has emerged as a eld of intense research
in the past decades. Dierent numerical methods were used to investigate crack formation in geological
structures [8, 9, 10]. In this work, we present a multiphase-eld simulation study of microfracturing in
sandstone, which takes an anistropic fracture toughness and lower crack resistance along the grain bound-
aries into account. The present model builts upon the framework of Schneider et al. [6] and ensures that
the mechanical jump conditions of the solid phases are fullled [5].
1 Mulitphase-eld model for crack propagation
We consider a computational domain consisting of N solid phases and a crack phase. Each phase is
represented by an order parameter φα (x, t), α ∈ [1, . . . , N, c], which describes the volume fraction of a
particular phase at position x and time t. The transition region between dierent phases is described by a
diuse interface in which the order parameter φα (x, t) varies continuously from 1 inside the phase α to 0
outside. The crack phase, represented by φc , describes the material degradation, where φc = 1 represents
a fully damaged and φc = 0 a fully intact material. At each point in the computational domain, the
∗ Corresponding author: michael.spaeth@kit.edu

†Authors contributed equally
381
PN
summation constraint of all occurring phase-elds α φα = 1 − φc is fullled. All the phase-elds are
collected in a phase-eld vector φ(x, t) = {φ1 (x, t), . . . φN (x, t), φc (x, t)} = {φs , φc }.
Based on the theory of Grith [2], a total free energy functional F to describe crack propagation in a
multiphase system of volume V is applied
Z
Gc (φ, ∇φc ) 1 2
F(φ, ∇φc , εel ) = ωc (φc ) + |∇φc | + fel (εel , φ) dV. (1)
V 2
| {z }
f (φ,∇φc ,εel )
The eective crack resistance is denoted by Gc (φ, ∇φc ) and fel (εel , φ) represents the eective strain
energy density. ∇φc is the gradient of crack phase-eld φc and εel is the elastic strain tensor according
to the innitesimal strain assumption. The crack resistance maps the energy of a crack in a regularised
manner by using a one-side well-potential ωc (φc ) = φ2c and a gradient term |∇φc |2 . The length scale
parameter controls the crack interface width. The evolution of the crack phase-eld is described in
terms of a variational derivative of the free energy functional which is given as

∂φc (x, t) 1 δF(φ, ∇φc , εel ) 1 ∂f (φ, ∇φc , εel )
∂f (φ, ∇φc , εel )
= −M = −M −∇· , (2)
∂t δφc ∂φc ∂∇φc
α
PN
where M is the mobility. With the interpolation function of solid phases hs (φ) = φα / β φβ the evolution
of solid phases reads
∂φα ∂φc
= −hα
s
(φ) for α = 1, .., N. (3)
∂t ∂t
This multiphase-eld model for crack formation can be extended to dierent applications by modifying the
crack resistance or strain energy density. For modeling crack propagation in brittle geological materials
a formulation for an anisotropic crack resistance is introduced in [4], where the crack resistance can be
modied along the c- and a-axis of arbitrary orientated crystals. Moreover for modeling intergranular
fracturing an interfacial crack resistance is presented. The crack resistance along grain boundaries can
be adapted with a third order polynomial function. An extension to elasto-plastic crack propagation
in geological materials with Drucker Prager plasticity is presented in [7]. Furthermore the presented
muliphase-eld model is extended in [3] for ductile fracturing in metals and is applied to a cast-iron
microstructure.
The static impulse balance ∇ · σ̄ = 0 is solved for the displacement eld with the degraded volume-
averaged stresses
N
X
∂ 1 α α α
σ̄ = (1 − φc )2
σ α
with
α
σ = ε · C [εel ] = C α [εα ], (4)
α=1
∂εαel
2 el el
where σα is the phase-inherent stress tensor, Cα is the phase-inherent stiness tensor, and εα
el
is the
phase-inherent elastic strain tensor. In multiphase regions the mechanical jump conditions are fullled
for the solid phases [5].
2 Numerical example
The capabilities of the presented model are demonstrated in a multiphase system containing quartz
grains. The simulation setup, the boundary conditions, and the crystallographic orientation of the grains
are depicted in Figure 1 a). In Figure 1 b) the crack resistance eld is illustrated. The crack resistance
along the a-axis is reduced based on the experimental values of the fracture toughness [1]. Furthermore
the crack resistance between grain boundaries is decreased. Schematically, the interfacial crack resistance
along the line (I-II) and the anisotropic crack resistance at the crack tip is visualised. Figure 1 c) depicts
the simulated crack path exhibiting both anisotropic transgranular and intergranular fracturing, resulting
in an irregular crack geometry.
Acknowledgments
The research has been supported by the German Research foundation (DFG) through the project
NE 822/34-1. Furthermore the authors thank the Helmholtz Association for the nancial support through
the program Renewable Energies (RE), ecient use of geothermal energy 35.14.01.
382
a) b) c)
intergranular
anisotropic crack growth
transgranular
crack growth
I II
I II
0 φc 1 0 Gc/GcMax 1
Figure 1. Crack propagation in a multigrain system, analogous to quartz sandstone. a) Simulation setup illus-
trating a microstructure with quartz grains along with applied boundary conditions and a pre-existing
crack. The crystallographic orientation of dierent grains is depicted schematically within each grain.
b) Crack resistance eld at the beginning of the simulation. The anisotropic crack resistance at the crack
tip and the interfacial crack resistance along path (I-II) is highlighted. c) Fully fractured specimen.
References
[1] Colin C Ferguson, Georey E Lloyd, and Robert J Knipe. Fracture mechanics and deformation
processes in natural quartz: a combined Vickers indentation, SEM, and TEM study. Canadian
Journal of Earth Sciences 24.3 (1987), pp. 544555.
[2] Alan Arnold Grith. VI. The phenomena of rupture and ow in solids. Philosophical transactions
of the royal society of london. Series A, containing papers of a mathematical or physical character
221.582-593 (1921), pp. 163198.
[3] Christoph Herrmann, Daniel Schneider, Ephraim Scho, Felix Schwab, and Britta Nestler. Multiphase-
eld model for the simulation of brittle and ductile crack propagation in grey cast iron microstruc-
tures (in preparation).
[4] Nishant Prajapati, Christoph Herrmann, Michael Späth, Daniel Schneider, Michael Selzer, and
Britta Nestler. Brittle Anisotropic Fracture Propagation in Quartz Sandstone: Insights from Phase-
eld Simulations (in preparation).
[5] Daniel Schneider, Oleg Tschukin, Abhik Choudhury, Michael Selzer, Thomas Böhlke, and Britta
Nestler. Phase-eld elasticity model based on mechanical jump conditions. Computational Mechan-
ics 55.5 (2015), pp. 887901.
[6] Daniel Schneider, Ephraim Schoof, Yunfei Huang, Michael Selzer, and Britta Nestler. Phase-eld
modeling of crack propagation in multiphase systems. Computer Methods in Applied Mechanics and
Engineering 312 (2016), pp. 186195.
[7] Michael Späth, Christoph Herrmann, Nishant Prajapati, Daniel Schneider, Felix Schwab, Michael
Selzer, and Britta Nestler. Multiphase-eld modelling of crack propagation in geological materials
with Drucker Prager plasticity (in preparation).
[8] Simon Virgo, Steen Abe, and Janos L Urai. Extension fracture propagation in rocks with veins:
Insight into the crack-seal process using Discrete Element Method modeling. Journal of Geophysical
Research: Solid Earth 118.10 (2013), pp. 52365251.
[9] XiaoLong Wang, Fang Shi, Chuang Liu, DeTang Lu, He Liu, and HengAn Wu. Extended nite
element simulation of fracture network propagation in formation containing frictional and cemented
natural fractures. Journal of Natural Gas Science and Engineering 50 (2018), pp. 309324.
[10] Xi Zhang and Robert G Jerey. Role of overpressurized uid and uid-driven fractures in forming
fracture networks. Journal of Geochemical Exploration 144 (2014), pp. 194207.
383
384
MS 20: Numerical methods in biomechanics

User-defined material subroutines for FEM calculations

based on automatic differentiation of strain energy func-
tions
1
Markus Hillgar̈tner ∗
, Theron Guo1 and Mikhail Itskov1
1 Department of Continuum Mechanics, RWTH Aachen University, Kackertstr. 9, 52072 Aachen, Germany
Abstract. The implementation of material models into the implicit finite element method (FEM) codes
requires the calculation of a fourth order tangent tensor. In hyperelastic models this tensor can be obtained
from the second derivative of the strain energy function with respect to the deformation gradient. Deriving
and implementing this derivative manually can be tedious and error prone.
This contribution presents an easy-to-use alternative based on automatic differentiation (AD) capable of
efficiently evaluating accurate derivatives [3, 6]. This technique is applied to automatically derive a user-defined
material subroutine (UMAT) for implicit FEM analysis of a hyperelastic anisotropic model of arterial layers
[1] in ABAQUS. Additionally, a pseudo-elastic model for the Mullins effect in filled rubber [7] is implemented
to evaluate the applicability of the discussed approach in cases where the material model exhibits inelastic
behavior described by internal variables.
1 Introduction
The implementation of material models into implicit finite element method (FEM) codes requires the
calculation of a fourth order tangent tensor. In constitutive models given in terms of a strain energy
function, this tensor can be obtained from the second derivative of tis function with respect to the
deformation gradient. The manual derivation and implementation of this derivative can be tedious and
error prone. In order to automatize this process, automatic differentiation of code can be a suitable tool
to create user defined material models without the need of manual differentiation. While the concept
has been used in the FEM context before [i.e. 10], we focus on its applicability inside the framework of
commercial FEM programs like ABAQUS.
2 Materials and methods
2.1 Constitutive modeling
Let us consider a material model using a multiplicative split of the deformation gradient F into an
isochoric (volume preserving) part F̄ and a volumetric (volume changing) part F̂ as
F = F̂F̄, J = det F, F̂ = J 1/3 I, F̄ = J −1/3 F, det F̄ = 1. (1)
Accordingly, the right Cauchy-Green tensor C and its isochoric part C̄ take the form
T
C = FT F, C̄ = F̄ F̄ = J −2/3 C, (2)
which leads to the invariants of C̄ as

h i
Ī = IC̄ = tr C̄, ¯ = II = 1 (tr C̄)2 − tr C̄2 ,
II ¯ = III = det C̄.
III (3)
C̄ C̄
2
∗ Corresponding author: hillgaertner@km.rwth-aachen.de ORCID iD: 0000-0001-7974-2055
385
In terms of a strain energy function Ψ the first and second Piola-Kirchhoff stress tensors P and S, and
the Cauchy stress tensor σ can be expressed as
∂Ψ ∂Ψ 1 1 ∂Ψ T
P= , S=2 , σ= PFT = F . (4)
∂F ∂C J J ∂F
In order to obtain convergence during a FEM iteration step, the calculation of a tangent tensor is
necessary. A variety of different tangent tensors exist:
The material tangent tensor (or Lagrangian tangent tensor) is given as the derivative of the second Piola-
Kirchhoff stress tensor S with respect to the Green-Lagrange strain E, which is its work conjugate strain
[4, 8]
∂S ∂S ∂2Ψ
C= =2 =4 . (5)
∂E ∂C ∂C∂C
The spatial tangent tensor c (or Eulerian tangent tensor) can be obtained from C using a push-forward
operation given as
1
cijkl = CIJKL FiI FjJ FkK FlL . (6)
J
In addition to the tangent tensors in the spatial or material formulation, one can define a mixed tangent
tensor using the first Piola-Kirchhoff stress P as [4]
∂P ∂2Ψ
A= = . (7)
∂F ∂F∂F
A is particularly useful in the context of automatic differentiation, as it utilizes the strain energy function
of interest Ψ and the deformation gradient F, which is the primary input in a user-defined subroutine for
a user material in the FE framework. The spatial tangent tensor c can be directly calculated for a given
A.
Differentiating Ψ with respect to the deformation gradient F is general, as it includes material models
based on the right Cauchy-Green tensor C = FT F or its invariants without additional effort. In less
general cases derivatives directly with respect to C or its invariants can be numerically favorable. In this
case, less general user subroutines such as UHYPER can be more suitable.
2.2 Automatic differentiation
The desired first (stress) and second (tangent) derivatives of the strain energy can be obtained in many
ways: While a handcoded analytical derivative is exact and often the numerically fastest method, creating
this code is time consuming and error prone. Subsequent minor changes in Ψ can lead to a huge amount of
outdated code, which needs to be updated manually. Alternatively, finite differentiation can be utilized.
This procedure, while easy to implement, only provides approximated results and often requires an
extensive amount of computational time. Automatic differentiation on the other hand guarantees exact
results at a computational effort. [5]
We choose the tangent mode of automatic differentiation, which applies the chain rule in forward direction.
To this end, Tapenade [3] is utilized twice: Once in order to obtain P = ∂Ψ ∂F and once again to obtain
∂2Ψ
A = ∂F∂F = ∂F . The specific stress and tangent tensors, required by the chosen FEM programm can
∂P
now be obtained based on P and A.
2.3 Implementation into an Abaqus/Standard user-defined material subroutine (UMAT)
The minimally required outputs of a UMAT subroutine in Abaqus/Standard are STRESS (Cauchy stress
σ) and the Jacobian matrix of the constitutive model DDSDDE (cAbaqus ). This tangent required by the
UMAT relates the volume-normalized Jaumann rate of the Kirchhoff stress to the symmetric rate of
deformation tensor [4, 8] and is given by
cAbaqus = c + c0 , (8)
386
where the fourth order tensor c0 can be expressed by Equation (6) as [4, 8]
1
c0ijkl = (δik σ jl + δil σ jk + δjk σ il + δjl σ ik ) , (9)
2
which possesses major and minor symmetries. δij denotes the Kronecker delta defined as
(
1 for i = j
δij = (10)
0 for i 6= j.
While c needs to be calculated for each material model, the implementation of c0 is independent of the
chosen material description.
In this study, three different material models have been implemented into UMATs: The neo-Hookean
strain energy (subsection 3.1) as the simplest case of an isotropic elastic model, the Ogden-Roxburgh
model (subsection 3.2) as extension of the neo-Hookean strain energy function to include the Mullins
effect and a model of arterial layers incorporating anisotropy (subsection 3.3). The models have been
implemented in terms of strain energy functions and the corresponding derivatives for the stress and the
tangent have been obtained using automatic differentiation.
3.1 Neo-Hookean strain energy
The neo-Hookean [9] hyperelastic material model can be represented as a strain energy function
ΨNH (Ī, J) = C10 (Ī − 3) + U (J) (11)
where C10 denotes a material constant and U (J) enforces near-incompressibility [2].
3.2 Ogden-Roxburgh pseudo-elastic model
The Ogden-Roxburgh model [7] extends the neo-Hookean strain energy function, given in equation (11),
to include the Mullins effect. This is achieved by scaling the elastic Cauchy stress σ 0 with a softening
function η(Ψ, Ψmax ), which depends on the current strain energy Ψ and the maximum strain energy
Ψmax in the load history. Accordingly,
σ OR = η(Ψ, Ψmax )σ 0 . (12)
This pseudo-elastic model utilizes an internal variable, implemented as solution-dependent state vari-
able STATEV, which needs to be correctly used and properly updated in the automatically differentiated
subroutine.
3.3 Fiber reinforced hyperelastic model
A simple hyperelastic model of arterial layers [1] utilizes the neo-Hookean strain energy presented in
subsection 3.1 for the fiber-reinforced ground substance. The n fiber families are modeled using
k1 h 2 i
Ψfi (C̄) = exp k2 Ēi (C̄) −1 , i = 1, ..., n, (13)
2k2
where k1 and k2 denote material parameters and Ēi represent strain-like quantities in the preferred
directions of fibers (please refer to [1] for details). The complete strain energy for the composite can be
given as
n
X
ΨG (F̄) = ΨNH (Ī, J) + Ψfi (C̄), I¯ = I(
¯ C̄), C̄ = C̄(F̄), J = J(F̄). (14)
i=1
387
4 Results and conclusion
All three investigated models have been successfully implemented in terms of the corresponding strain
energy Ψ and then derived automatically in order to create fully functional UMAT subroutines. Since the
applied automatic differentiation procedure grants exact derivatives, the results and convergence behavior
did not differ significantly from the manually programmed ones.
While the analysis of computational times for the different examples is still in progress, the difference in
implementation time in comparison to manual programming is enormous. Once a template is created,
a strain energy function can be automatically turned into a subroutine without manual differentiation.
This advantage in the implementation can save a considerable amount of time during the development
and testing phases of new constitutive models.
References
[1] Thomas Christian Gasser, Ray W. Ogden, and Gerhard A. Holzapfel. Hyperelastic modelling of
arterial layers with distributed collagen fibre orientations. Journal of the Royal Society, Interface
3.6 (2006), pp. 15–35.
[2] Stefan Hartmann and Patrizio Neff. Polyconvexity of generalized polynomial-type hyperelastic
strain energy functions for near-incompressibility. International Journal of Solids and Structures
40.11 (2003), pp. 2767–2791.
[3] Laurent Hascoet and Valérie Pascual. The Tapenade automatic differentiation tool: Principles,
model, and specification. ACM Transactions on Mathematical Software 39.3 (2013), pp. 1–43.
[4] Maria A. Holland. The Hitchhiker’s Guide to Abaqus. 2017.
[5] Charles C. Margossian. A Review of automatic differentiation and its efficient implementation.
January (2018), pp. 1–19.
[6] Uwe Naumann. The Art of differentiating computer programs: an introduction to algorithmic dif-
ferentiation. Society for Industrial and Applied Mathematics, 2012. isbn: 161197206X.
[7] Ray W. Ogden and David George Roxburgh. A pseudo–elastic model for the Mullins effect in
filled rubber. Proceedings of the Royal Society of London. Series A: Mathematical, Physical and
Engineering Sciences 455.1988 (1999), pp. 2861–2877.
[8] V. Prot and B. Skallerud. Nonlinear solid finite element analysis of mitral valves with heterogeneous
leaflet layers. Computational Mechanics 43.3 (2009), pp. 353–368.
[9] R. S. Rivlin. Large Elastic Deformations of Isotropic Materials. IV. Further Developments of the
General Theory. Philosophical Transactions of the Royal Society A: Mathematical, Physical and
Engineering Sciences 241.835 (1948), pp. 379–397.
[10] Florian Zwicke, Philipp Knechtges, Marek Behr, and Stefanie Elgeti. Automatic implementation
of material laws: Jacobian calculation in a finite element code with TAPENADE. Computers &
Mathematics with Applications 72.11 (2016), pp. 2808–2822.
388

Nonlinear fibre distribution in a finite element model of

a human left ventricle for sparse and non-constant data
sets
1
David Holz ∗ , Minh Tuan Doung1,3 , Myoungsic Kim1 , Muhannad Alkassar2 , Sven Dittrich2 and Sigrid
Leyendecker1
1 Chair of Applied Dynamics, University of Erlangen-Nuremberg
2 Pediatric Cardiology, University of Erlangen-Nuremberg
3 School of Mechanical Engineering, Hanoi University of Science and Technology
Abstract. The morphology of the cardiac tissue is one of the most important characteristics to ensure a
proper and choreographed contraction and thus the overall pump function of the heart. The morphology,
especially the characteristic fibre structure of the heart, substantially influences the electrical and mechanical
function. Thus, the correct representation of the specific fibre orientation is fundamental for all computa-
tional models in cardiac research. Often, there is a lack in experimental data or DT-MRI and thus rule-based
schemes have to be applied in order to capture the realistic and patient specific fibre orientation as accu-
rate as possible. For this purpose, we propose a finite element based nonlinear fibre interpolation scheme.
Introduction
The specific morphology and especially the fibre orienation of the cardiac tissue is the basis for the
characteristic motion and pump function of the heart. Thus, it is fundamental in cardiac modelling to
realistically represent the fibre orientation since it substantially influences the mechanical and electro-
physiological properties. The fibre orientation of the cardiac tissue can be measured e.g. using DT-MRI
or histology. However, in the case that no patient-specific fibre orientation is known or only sparse data
from clinical measurements is available, rule-based models are applied. Under these circumstances, it
is assumed that the fibre angles are measured at discrete points of the heart, often on the endo- and
epicardium (e.g. from simplified reference measurements on the surface rather than in the volume or lay-
ers). Additional information about the transmural change in fibre angle are obtained e.g. from histology
slides. Here, however, we assume that a nonlinear function representing the transmural change in fibre
angle is given, see Figure 1.
60
m=1
m=0.5
40 m=2
m=3
20
fibre angle α (°)
-20
-40
-60
0 0.2 0.4 0.6 0.8 1
endo transmural depth e epi
Figure 1. Left: transmural fibre angle α(e) = R ∗ sgn(1 − 2e)|1 − 2e|m depending on the transmural depth
e ∈ [0, 1], R =60◦ represents the maximum fibre angle, m ∈ N is to be chosen in accordance with clinical
measurements, see [1]. Right: characteristic morphology of the cardiac tissue with the transmurally
varying fibre direction f , sheet direction s and normal direction n.
∗ Corresponding author: david.holz@fau.de ORCID iD: https://orcid.org/0000-0003-4838-718X
389
Based on the approach by Wong [3], we propose a finite element-based approach to assign the smooth
nonlinear fibre distribution in a human left ventricle. Therefore, we solve an anisotropic Laplace equation,
with a scalar-valued unknown, representing a specific vector component of the fibre direction f ∈ R3 .
We show that the proposed interpolation scheme is capable of creating a smooth nonlinear interpolation
for a geometrically complex patient-specific left ventricle obtained from high-resolution MRI at the Uni-
versitätsklinikum Erlangen-Nürnberg. Often sparse and insufficient data sets are available as boundary
conditions for the Laplace problem. Analytical or least squares approaches can usually not be used or
only under certain restrictions. However, also in this case, the proposed method provides an easy and
robust way to create a smooth nonlinear fibre distribution with acceptable errors in fibre angle.
We solve an anisotropic Laplace-equation for each vector componentfi with i = x, y, z of the fibre vector
f.
∇ · (λ∇fi ) = 0 in E (1)
with given Dirichlet and Neumann boundary conditions (b.c.)
fi = f¯i on ∂E, (2)
The partial differential Equation (1) is transformed into its weak form and integrated over the domain,
being discretised with linear tetrahedral elements. We define the Dirichlet b.c. as well as the element
specific diffusion coefficient λi,n in the sequel.
1.1 Dirichlet boundary conditions
To define the Dirichlet boundary conditions, we assign the given (in an ideal case measured) fibre orien-
tation (more specifically absolute values of vector components) to local coordinate frames on the nodes of
the finite element mesh on the boundary, see Figure 2. Here, we assume that the fibre orientation on the
entire endo- and epicardium is given, being +60◦ and -60◦ with respect to the circumferential direction,
respectively.
Figure 2. Left: characteristic microstructure of the cardiac tissue represented by local coordinate frames at each
finite elemet node. Following [2], the cardiac tissue is characterized by an orthotropic structure con-
sisting of the fibre direction f , sheet direction s and normal direction n. Right: visualisation of the
fibre orientation in layers of the heart. The red colour represents the positive angle of +60◦ on the
endocardium, the blue colour represents the negative angle of -60◦ on the epicardium with respect to
the circumferential direction.
For more detailed information about the definition of the local coordinate system, see [3].
390
1.2 Diffusion coefficient
To assign the element and component specific diffusion coefficient λi,n , where i = x, y, z and n accounts
for the specific element, we consider the influence of the following three factors: vector component
interpolation (1), area dependency (2) and nonlinear fibre distribution (3).
(1) vector component interpolation

1 60
0.8
40 linear
0.6 calculated angle
0.4
20
fibre angle α (°)

0.2
Fy
0 0
-0.2
-20
-0.4
-0.6
-40
-0.8
-1 -60
-0.4 -0.2 0 0.2 0.4 0.6 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fx transmural depth e
Figure 3. Vector component interpolation for a simple cube with α = +60◦ on endo- and α = −60◦ on epicardium
(left) leads to a nonlinear angle interpolation (right), even though the prescribed transmural fibre angle
α(e) is linear (m = 1). Based on the spherical linear interpolation, we propose Equation (3) to take
into account the nonlinear change of the vector components, where en is is the transmural depth of the
specific element n, Ai,n and Ci,n are the vector components of the tangent vectors (of the fibre direction
f ) on the endo- and epicardium and B defines the angle spanned by An and Cn .
(1) 1 (Ai,n · sin((1 − en ) · B) + Ci,n · sin(en · B)

λi,n = with wi,n (en , Ai,n , Ci,n ) = (3)
|wi,n | sin(B)
(2) area dependency

60
40
20
fibre angle α (°)
-20
-40
-60
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
transmural depth e
60
fibre angle
40 m=1
20
fibre angle α (°)
-20
-40
-60
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
transmural depth e
Figure 4. Top left: transmural fibre angle α without accounting for the different area on the endo-and epicardium
(shift of 0◦ line to endo; e < 0.5). Top center: transmural fibre angle α taking into account the area
change (α = 0◦ at e = 0.5). Top right: fibre angle α over the transmural depth e without accounting
for the area. Bottom - left and center: comparison of the transmural depth computed by a diffusion
problem with a constant isotropic diffusion coefficient (Dirichlet b.c.: 0 on endo and 1 on epi; eth is the
transmural depth for each element) and the transmural depth ecal calculated by following the steepest
descent of the diffusion field for an MRI-based left ventricle. Bottom - right: linear angle interpolation
following the proposed Equation (4) to account for the local area change in the LV.
e 2
ncal
λ(2)
n = (4)
enth
391
(3) nonlinear fibre distribution
The nonlinear fibre interpolation is also influenced by the derivative of the predefined function α(en ).
1
λ(3)
n = (5)
T 0 (en )
(tot)
Thus, the total description of the elementwise and vector component dependent diffusion coefficient λi,n
reads
(tot) (1)
λi,n = λi,n · λ(2) (3)
n · λn . (6)
2 Results
In Figure 5 (left), the nonlinear fibre distribution with m = 3 for a left ventricle is shown. The proposed
method creates a smoothly varying interpolation with acceptable errors in angle interpolation.
60
16
40 14
12
20
angle error (%)
fibre angle α (°)
10
0
8
-20 6
4
-40
2
-60
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0
20 30 40 50 60 70 80 90 100
transmural depth e ratio of surface sampling (%)
Figure 5. Left: fibre angle α over the transmural depth for each element (fibre vectors in the apex area reassigned).
Right: sensitivity analysis shows that even for a reduced set of Dirichlet boundary conditions (up to
20%), the method provides results with acceptable errors (max. error approx. 15◦ ).
Figure 5 (right) shows that even in the case of sparse data, represented by a reduced set of Dirichlet
boundary conditions (up to 20%), the method provides results with acceptable errors (max. error ap-
prox. 15◦ ).
In this work, we show that the proposed method is capable of creating a smooth nonlinear fibre interpo-
lation. The method is easy to apply, robust and even employable in the case of sparse data. Further, the
method can be extended to also take into account non-constant fibre angles on the endo- and epicardium.
Acknowledgments
The authors thank the Klaus Tschira Stiftung grant 00.289.2016 for funding support
References
[1] V. Carapella, R. Bordas, P. Pathmanathan, M. Lohezic, J. Schneider, P. Kohl, K. Burrage, and V.
Grau. Quantitative Study of the Effect of Tissue Microstructure on Contraction in a Computational
Model of Rat Left Ventricle. PLOS ONE 9 (2014), pp. 1–12.
[2] G. A. Holzapfel and R. W. Ogden. Constitutive modelling of passive myocardium: a structurally
based framework for material characterization. Royal Society of London Philosophical Transactions
A: Mathematical, Physical and Engineering Sciences 367 (2009), pp. 3445–3475.
[3] J. Wong and E. Kuhl. Generating fibre orientation maps in human heart models using Poisson
interpolation. Computer methods in biomechanics and biomedical engineering 17 (2014), pp. 1217–
1226.
392

A Frictional and Adhesive Contact Model for

Debonding of the Bone-Implant interface Based on
State Variable Friction Laws
∗ 1,2 1 2 2 1
Katharina Immel , Thang X. Duong , Vu-Hieu Nguyen , Guillaume Haïat and Roger A. Sauer
1 Aachen Institute for Advanced Study in Computational Engineering Science (AICES), RWTH Aachen University, Templergraben
55, 52056 Aachen, Germany

2 CNRS, Laboratoire Modélisation et Simulation Multi Echelle, MSME UMR 8208 CNRS, 61 Avenue du Général de Gaulle, 94010
Créteil Cedex, France
Abstract. Implant loosening is one of the major problems of cementless orthopedic and dental implants.
However, the underlying phenomena of osseointegration, especially regarding frictional and adhesive eects at
the interface and their inuence on long term stability, remain poorly understood. Most numerical studies so far
assume either a fully bonded interface, frictionless contact or pure Coulomb's friction. While this assumption
might be valid for the prediction of initial stability, the prediction of failure of osseointegrated implants requires
the inclusion of more complex frictional and adhesive eects. Here, a macroscopic contact formulation is
proposed based on state variable friction laws [5] that can model debonding at the bone-implant interface. For
each element on the contact surface, the tangential contact stress depends on the normal contact pressure, the
sliding rate, and the current state of the element (fully bonded, debonding or fully sliding) with respect to its
sliding history and initial bonding state. This model leads to dierent contact laws for every state, for example,
a frictional adhesion model [4] for fully bonded elements and pure Coulomb's friction for fully sliding once. This
contact formulation is embedded in a nite element contact framework [2], including a NURBS-enrichment
approach for three-dimensional contact elements [1] for an ecient contact treatment. It is validated by
comparing the numerical results with experimental data [3] and used to model the torsional debonding of
osseointegrated implants. The results show good agreement with experimental data and provide estimates
for the material properties of the bone-implant interface, such as the shear modulus, Young's modulus, and
friction coecient. Furthermore, it allows the analysis of the debonding behavior, for example, in terms of
maximum bearing loads and stresses, and crack propagation, that are dicult to measure in vivo or in vitro.
Introduction
Cementless, endosseous implants are widely used in orthopedic and oral surgery. These implants achieve
primary stability during surgery by mechanical interlocking, while long term stability is achieved by bone
growing around and into the porous or rough surface of the implant, a process called osseointegration.
However, failures still occur and may cause the debonding of the bone-implant interface, which in turn
leads to a revision of the implant. It remains dicult to predict implant failure since the underlying
osseointegration phenomena, especially in terms of friction and adhesion, and their inuence on the
primary and secondary stability of the implant, is still poorly understood. We therefore propose a
contact formulation that can model the debonding behavior of osseointegrated implants.
Given two 3D bodies B1 and B2 and their contact surface ∂c B , the contact traction tc can be decomposed
into a normal and tangential component, i.e. tc = tn − tt . The magnitude of the normal traction is equal
to the normal contact pressure, i.e.
ktn k = p, (1)
∗ Corresponding author: immel@aices.rwth-aachen.de
393
and is always compressive if no adhesion is present. For frictional contact, the tangential traction is
determined by the behavior during sticking and sliding. The distinction between these two states is
based on a slip criterion of the form
(
< 0, for sticking,
fs (2)
= 0, for sliding.
It can be formulated as
fs = ktt k − tmax
t
, ktt k ≤ tmax
t
, (3)
where tmax
t
> 0 is the maximal tangential contact traction. For sticking, the tangential traction is dened
by the constraint that no relative tangential motion occurs, while for sliding, the tangential traction is
dened by a sliding law. Most numerical analyses of implant stability only consider either fully bonded
contact, frictionless contact or pure Coulomb's friction, according to
ġ s
tt = −µ p , (4)
kġ s k
where µ is a constant friction coecient, p is the normal contact pressure, and ġ s denotes the sliding rate.
Such an approach may be a valid assumption for the assessment of initial stability, where no adhesive
forces are expected. However, to predict the failure of osseointegrated implants, adhesive and frictional
eects need to be taken into account. One approach to account for the change of contact behavior due
to dierent states of the interface are state variable friction laws, introduced by Rice and Ruina [5]. In
general, these laws assume that at any given time, the frictional stress of a point x on the contact surface
depends on the normal pressure, the sliding rate ġs = kġ s k and a state variable S , i.e.
tt = tt (p, ġs , S). (5)
Although these laws are mainly applied in geology and geophysics, the state variable S can be interpreted
as the current bonding state of a certain point x on the bone-implant interface ∂c B . Here, S is modeled
as a smooth step function from 1 to 0 and depends on the total displacement of the contact surface. This
denition can be used to model macroscopic frictional and adhesive phenomena at the interface, e.g.
• adhesive contact while bonded (S > 0),
• debonding due to displacement (0 < S < 1),
• frictional sliding when fully debonded (S = 0).
The corresponding current state then determines the appropriate contact law. An example is using the
frictional adhesion model of Mergel et al. [4] for fully bonded and Coulomb's law for debonded points.
Adhesion in normal direction can be modeled, e.g. by an exponential cohesive zone model [7] or van der
Waals adhesion [4]. This contact formulation is implemented into an existing nite element (FE) contact
formulation, based on Duong and Sauer [2] and Sauer and De Lorenzis [6]. It uses an ecient contact
enrichment approach based on Non-Uniform Rational B-Splines (NURBS) basis functions for the contact
elements and Lagrangian basis functions for the bulk, see Corbett and Sauer [1].
2 Application
This contact model is used to simulate mode III debonding of three-dimensional, osseointegrated implants,
based on experimental data from Mathieu et al. [3]. The problem setup is illustrated in Figure 2. The
implant is meshed with 72 linear Lagrangian bulk elements and 36 quadratic NURBS-enriched contact
elements, while the bone block is meshed with 648 bulk elements and 324 contact elements. The mesh
and the contact surface of the implant are depicted in Figure 1. The lower surface of the bone block
is xed in all directions, while the sides are xed in their corresponding normal direction. The bone
material parameters are Gb = 8 GPa, Eb = 21 GPa, νb = 0.3, for the shear modulus, Young's modulus,
and Poisson ration, respectively. The friction coecients and the average contact pressure are µ1 = 0.4,
µ2 = 0.55 and p̄ = 3.1 MPa, and are determined by a parameter tting. The cylindrical titanium implant
(Ti-6Al-4V) is rotated by 10◦ around its central axis with a constant load step size of 0.01◦ /s. In the
following we analyze the resulting torque Mt as a function of the imposed rotation angle θ.
394
MT
Figure 1. NURBS-enriched nite element mesh. Left: complete geometry. Right: Enlargement of the contact
surface of the implant.
Figure 2 shows the torque-per-angle curve for one sample obtained by the experiments and the analytical
model of Mathieu et al. [3] compared to the FE results of the present contact model. It can be seen that
our proposed model results in an overall better t of the experimental measurements than the analytical
model. The analytical model was t to match the peak and the decrease in torque, which resulted in a
mean relative error of 11.26 %. The proposed contact model can reproduce more characteristics of the
torque curve, such as the initial slope, the peak, the softening, and the torque for complete debonding,
which leads to a mean relative error of 2.15 %.
Some minor deviations from the experimental results can be seen for θ > 2.4◦ where the decrease of
the torque is not exactly matched. A dierent transition function for S may allow for a closer t there.
Another deviation can be seen in between 0.2◦ and 1.8◦ where the FE solution shows an idealized peak.
Mathieu et al. [3] showed that part of the limitations of their model lies in the assumption of a fully bonded
interface at the beginning of the experiment. A visual analysis of the implants' surfaces after debonding
indicated that most likely, full osseointegration was not achieved. This resulted in regions where no bone
tissue was initially attached to the implant surface and thus no adhesive or frictional eects can take place
there. As the peak in the torque-per-angle curve represent the part of the deformation where adhesive
forces are expected to have a large inuence, considering an interface that is only partially bonded in the
beginning might improve the numerical results there.
0.06
0.05
torque MT [Nm]
0.04
0.03
0.02
experimental measurements from [3]
0.01 analytical solution from [3]
present FE solution
0
0 1 2 3 4 5 6 7 8 9 10
rotation angle θ [deg]
Figure 2. Mode III debonding of an implant. Variation of the torque MT as a function of the imposed rotation
angle θ.
395
3 Conclusion
This work presents a frictional and adhesive contact model, based on a state variable friction law and
its application to the debonding of osseointegrated implants. A numerical solution with the new friction
model is applied to the mode III debonding of a coin-shaped implant and the results are compared to
experimental and analytical results from the works of Mathieu et al. [3]. Overall it is shown, that assuming
a smooth transition from an unbroken to a broken state, characterized by a decreasing friction coecient
during the debonding process, leads to good agreement with experimental data. It also provides good
estimates for the interface properties, such as the shear modulus, Young's modulus, and friction coecient.
Furthermore, it enables the analysis of the debonding behavior, e.g. in terms of maximum bearing loads
and stresses and crack propagation. An extension to partial osseointegration, rough surfaces, as well as
realistic implant geometries, will be considered in future works.
Acknowledgments
This work has received funding from the European Research Council (ERC) under the European Union's
Horizon 2020 research and innovation program (grant agreement No 682001, project ERC Consolidator
Grant 2015 BoneImplant) and the Jülich Aachen Research Alliance Center for Simulation and Data
Science (JARA-CSD) School for Simulation and Data Science (SSD).
References
[1] Callum J. Corbett and Roger A. Sauer. Three-dimensional isogeometrically enriched nite elements
for frictional contact and mixed-mode debonding. Computer Methods in Applied Mechanics and
Engineering 284 (2015), pp. 781806.
[2] Thang X. Duong and Roger A. Sauer. A concise frictional contact formulation based on surface
potentials and isogeometric discretization. Computational Mechanics (2018), pp. 120. doi: https:
//doi.org/10.1007/s00466-019-01689-0.
[3] Vincent Mathieu, Romain Vayron, Etienne Barthel, Davy Dalmas, Emmanuel Soer, Fani Anag-
nostou, and Guillaume Haïat. Mode III cleavage of a coinshaped titanium implant in bone: Eect
of friction and crack propagation. Journal of the Mechanical Behavior of Biomedical Materials 8.1
(2012), pp. 194203.
[4] Janine C. Mergel, Riad Sahli, Julien Scheibert, and Roger A. Sauer. Continuum contact models for
coupled adhesion and friction. The Journal of Adhesion (2018), pp. 133. doi: 10.1080/00218464.
2018.1479258.
[5] James R. Rice and Andy L. Ruina. Stability of steady frictional slipping. Journal of Applied Me-
chanics50.2 (1983), pp. 343349.
[6] Roger A. Sauer. Local nite element enrichment strategies for 2D contact computations and a cor-
responding post-processing scheme. Computational Mechanics 52.2 (2013), pp. 301319.
[7] Roger A. Sauer and Laura De Lorenzis. A computational contact formulation based on surface
potentials. Computer Methods in Applied Mechanics and Engineering 253 (2013), pp. 369395.
396

Finite element technologies for stent simulations

∗1 2 1
Kiran Manjunatha , Jan Frischkorn and Stefanie Reese
1 Institute of Applied Mechanics, RWTH Aachen University, Mies-van-der-Rohe Str. 1, D-52074 Aachen, Germany
2 Volkswagen AG, Wolfsburg
Abstract. Solid-beam formulation (Q1STb) is a nite element technology developed with the aim of con-
ducting accurate simulations of beam-like structures (e.g. stents) with minimal element count. The re-
duced integration solid element formulation (Q1SP) with hourglass stabilization is an ecient and locking-
free alternative for standard brick elements and can handle incompressibility as well as a wider range of
aspect ratios. The scope of this work is to investigate the usage of a combination of both the aforemen-
tioned formulations in analyzing drug-eluting stents. It will be shown that this strategy is viable and e-
cient in comparison to the element formulations cuurently available in commercial nite element softwares.
Introduction
Stents are medical devices inserted into blood vessels to treat atheroscleroses and hence recover blood ow
to normality, and also to stabilize aneurysms. Most commercial stent congurations comprise of ligree
constructions which render the three-dimensional nite element stenting simulations computationally
expensive. Signicant expenses can be attributed to the following factors: Accurate resolution of local
phenomena (e.g. plasticity, phase transformations, damage etc.) requires multiple elements across the
cross-section of the stent struts. Also, to restrict the element aspect ratios within reasonable bounds, dense
discretizations in the longitudinal direction of stent struts are expected. Additionally, modeling material
nonlinearities, large deformations and contact incur signicant computational costs. Lastly, the presence
of multiple layers within the stent cross-sections (e.g. drug-eluting polymer layer, gold reinforcement layer
for welds etc.) exacerbates the cost of high-delity simulations.
Figure 1. Coronary stent
In the context of drug-eluting stents, development of a robust and computationally ecient strategy for
high-delity simulations is a challenging task. As a rst step in this regard, we investigate the use of a
combination of two advanced nite element formulations for eciently modeling the mechanics of stent
deployment. The solid-beam formulation [1] is utilized to model the metallic core. A reduced integration
solid element formulation with hourglass stabilization [2] is utilized to model the drug-eluting polymer
layer.
∗ Corresponding author: kiran.manjunatha@ifam.rwth-aachen.de
397
1 Finite element technologies
1.1 Solid-beam formulation (Q1STb)
The short-hand notation Q1STb denotes a formulation based on the classical "Q1" element, "ST" indi-
cating "stabilization based on Taylor expansion " and the "b" indicating the (solid-)beam type of element
within the so-called Q1STx family of elements. A brief outline of the formulation in its current state is
provided below.
The foundation of the formulation is a mixed variational principle based on the Hu-Washizu three-eld
functional with the displacement, the stress and the strain eld being treated as primary unknowns.
Curvature thickness locking and signicant portion of transverse shear locking are alleviated using the
assumed natural strain (ANS) method. Poisson thickness locking, which occurs due to the ANS-modied
transverse normal strains, and transverse shear locking in the beam plane are cured using the enhanced
assumed strain (EAS) method. Volumetric locking is partially treated by the EAS method.The reduced
integration concept, within which the formulation is embedded, helps eliminate volumetric locking com-
pletely. Full integration is performed on the transverse plane at the center of the beam axis and reduced
integration is carried with respect to the beam axis. Ecient analytical integration of hourglass parts of
the stiness matrix as well as the internal force vector is achieved by (a) approximating the inverse Jaco-
bian via a Taylor series expansion with respect to the element center. (b) reducing the number of terms
within the polynomial approximations of the Green-Lagrange strain and the B-operator via Taylor series
expansions with respect to the element center. (c) a Taylor series expansion of the second Piola-Kirchho
stress with respect to the transverse plane at the center of the beam axis. The adaptive hourglass material
tangent is dened by means of the deviatoric part of the linear elasticity matrix containing an eective
shear modulus term which can be interpreted as an average secant stiness at the quadrature points.
The reader is directed to [1] and [4] for implementation details.
1.2 Reduced integration solid element formulation with hourglass stabilization (Q1SP)
The short-hand notation Q1SP denotes a formulation based on the classical "Q1" element, "S" indicat-
ing a "stabilized" element and "P" indicating the assumption of an equivalent "parallelepiped". The
formulation is based on the non-linear Hu-Washizu principle and the stabilization concept considers the
full tangent (except in compression) as opposed to the earlier works wherein the geometrical stiness
was neglected to ensure stability. Volumetric and shear locking are alleviated by introducing a modied
stabilization stiness contribution which involves the assumption of an equivalent parallelepiped and the
enrichment of strains using the enhanced strain method introduced in [5] and [6]. The stabilization fac-
tors in the large deformation regime are computed by arriving at an expression analogous to the linear
elasticity case. The reader is directed to [2] for implementation details.
No special considerations are necessary for coupling the two formulations since they behave as solid
elements in this context. Also, the coupled setup can handle double-sided contacts which are often
encountered in stent simulations.
2 Numerical investigations
Based on the results presented in [3] , a critical section of the drug-eluting stent (Fig. 1), which repeats
itself throughout the structure, is chosen for our investigations. Cross-section with just the metal core as
well as that including the polymer layer are investigated under the boundary conditions shown below.
The parameters R = 0.2 mm, θ = 8.5◦ , h = 0.0675 mm for both type of cross-sections. For the bare-metal
strut, l = 0.912 mm and for the multi-layered strut, l = 0.92 mm. Within the cross-sections, ts = 0.8
mm and tp = 0.08 mm.
The sum of reaction forces at the point of application of displacement D = 0.25 mm is investigated and
presented.
398
θ h
D D k 3
R l
3
n m 1 2
1
(a) geometry
(b) boundary conditions
Figure 2. Stent strut
tp
ts
ts
tp
ts
(a) bare-metal
ts
(b) multi-layer
Figure 3. Strut cross-sections (k − l − m − n)
Table 1. Material parameters

Layer Material model E ν σy H
Metal core elasto-plasticity with linear isotropic hardening 200 GPa 0.3 370 MPa 7 GPa
Polymer Neo-Hooke 1 GPa 0.49 - -
Abbreviations: N QT = number of quadrature points on the trasnverse plane at the center of beam
axis, N EL = number of elements longitudinally along the radius, N ET = number of elements across the
thickness.
First, the bare-metal stent strut is analyzed with one element across the thickness for varying N EL and
N QT . Further, by xing N EL and N QT , the eect of varying N ET is analyzed and compared to the
results using C3D8I elements within ABAQUS. The sum of reaction forces F is normalized using the value
obtained by using a ne discretization of 8 C3D8I elements across the thickness. Finally, results for the
multi-layered strut using Q1STb and Q1SP are compared with the value obtained using a discretization
of 8 C3D8I elements across the thickness in the metal layer and 2 C3D8R elements across the thickness
in the polymer layer. For the multi-layered cross-section, the polymer layer is discretized with just one
element across the thickness when using Q1SP.
0.098
NQT = 3
0.096 NQT = 5
NQT = 7
0.094 NQT = 10
NQT = 12
0.092 NQT = 14
0.09
0.088
0.086
0.084
0.082
24 8 16 32 64
Figure 4. N EL and N QT convergence
399
1.12 1.12
Q1STb (NQT = 12) Q1STb (NQT = 12) + Q1SP
1.1 C3D8I 1.1 Reference - C3D8I + C3D8R
Reference - C3D8I
1.08 1.08
1.06 1.06
1.04 1.04
1.02 1.02
1 1
0.98 0.98
1 2 3 5 7 1 2 3
(a) bare-metal (b) multi-layer
Figure 5. N ET convergence
3 Conclusions
It is observed that with just one element across both the layers, results comparable to those using ne
discretizations of the strut with C3D8I and C3D8R elements are obtained. Hence, usage of Q1STb and
Q1SP formulations for stent deployment analyses is a viable and an ecient choice.
Acknowledgments
Financial support of the project 395712048 by the German Science Foundation (DFG) is gratefully
acknowledged.
References
[1] J. Frischkorn and S. Reese. A solid-beam nite element and non-linear constitutive modelling. Com-
put. Methods Appl. Mech. Engrg. 265 (2013), pp. 195212.
[2] S. Reese, P. Wriggers, and B.D. Reddy. A new locking-free brick element technique for large defor-
mation problems in elasticity. Computers and Structures 75 (2000), pp. 291304.
[3] A Schiavone, T Y Qiu, and L G Zhao. Crimping and deployment of metallic and polymeric stents -
nite element modelling. Vessel Plus 2017 Jan 19.[Online First] (2019).
[4] M. Schwarze and S. Reese. A reduced integration solid-shell nite element based on the EAS and the
ANS conceptLarge deformation problems. Int. J. Numer. Meth. Engng 85 (2011), pp. 289329.
[5] J. C. Simo and F. Armero. Geometrically non-linear enhanced strain mixed methods and the method
of incompatible modes. Int. J. Numer. Meth. Engng 33 (1990), pp. 14131449.
[6] J. C. Simo and M. S. Rifai. A class of mixed assumed strain methods and the method of incompatible
modes. Int. J. Numer. Meth. Engng 29 (1990), pp. 15951638.
400

Numerical simulation of the viral entry into a cell by

receptor driven endocytosis
T. Wiegold∗1 , S. Klinge1 , R. P. Gilbert2 and G. A. Holzapfel3,4
1 TU Dortmund University, Institute of Mechanics, Leonhard-Euler-Str. 5, 44227 Dortmund, Germany
2 Universityof Delaware, Department of Mathematical Sciences, 312 Ewing Hall, Newark, Delaware 19716, USA
3 Graz University of Technology, Institute of Biomechanics, Stremayrgasse 16/2, 8010 Graz, Austria
4 Norwegian University of Science and Technology (NTNU), Department of Structural Engineering, 7491 Trondheim, Norway
Abstract. The present contribution deals with the viral entry into a cell dictated by the change of the receptor
density. While the receptors of the virus are considered to be fixed on the surface, the ones of the cell
are mobile over the membrane. Upon contact, a locally increased receptor density of the cell is required,
which causes the receptors to diffuse towards the virus. In the contact (adhesion) zone the membrane inflicts
and forms an envelope around the virus. The vesicle created in this process imports its cargo into the
cell. The described problem is simulated based on the assumption that the differential equation typical of
the heat transport together with two boundary conditions is suitable to describe the diffusion of receptors.
The differential equation is accompanied by two boundary conditions. The first condition considers the
balance of fluxes on the adhesion front, where the amount of receptors required for advancement has to
be equal to the amount provided by the cell. At the front, the velocity is supposed to be proportional to
the gradient of the chemical potential. The second condition considers the energetic aspects of the process.
Here, contributions due to the binding of receptors, the free energy of the membrane, its curvature and
the kinetic energy due to the motion of the front are taken into account. The resulting moving boundary
problem consisting of the differential equation and the two boundary conditions is well posed. The finite
difference method as a direct numerical method for the solution of the differential equation is used to solve
the underlying system. Numerical examples investigate the influence of the mobility of receptors as well as
of different starting configurations in the cells and virus receptor distribution on the endocytosis process.
Introduction
One of the main mechanisms used by viruses in order to enter a cell is receptor driven endocytosis
[gibbons2010]. During this process the virus connects its receptors to the receptors of the cell. This
causes the cell membrane to bend and to build a vesicle around the virus. The process is characterized
by the ability of the receptors of the cell to move freely within its membrane, while the receptors of
the virus are fixed and occur with a much higher density. Upon contact, the virus dictates the required
receptor density for bonding, thus causing the receptors of the cell to move towards the contact area. As
a consequence, the contact area grows while the local receptor density of the cell changes, resulting in a
moving boundary problem.
1 Endocytosis as a moving boundary problem
At the beginning of the process a virus connects to the cell and splits the cell membrane into two areas
separated by an adhesion front. These two areas are characterized by the distribution of the receptor
density ξ(x, t). In the contact area the cell receptor density matches the constant receptor density of the
virus ξ(x, t) = ξeq . In the area in front of the adhesion zone the density is a function of position and
time ξ(x, t) with its minimum at the adhesion front, where ξ(x, t) = ξ+ . Since the cell size is orders of
magnitudes larger than the virus size, the density profile remains unchanged far away from the contact
∗ Corresponding author: tillmann.wiegold@tu-dortmund.de
401
area lim ξ(x, t) = ξ0 . The motion of the receptors on the cell surface is based on the diffusion differential
x→∞
equation depending on the mobility m, i.e.
∂ξ ∂2ξ
= m 2. (1)
∂t ∂x
A solution of this partial differential equation requires additional boundary conditions. The conservation
of binders is considered first. It states that the amount of receptors required for the advancement of the
adhesion front has to be equal to the amount of receptors provided by the flux. Following Fick’s first
law, the diffusion equation leads to the first boundary condition

∂ξ
(ξeq − ξ+ ) v+ − m = 0, (2)
∂x +
which depends on quantities related to the adhesion front [freund2004]. Here, v+ denotes the velocity
of the front. For the second boundary condition an approach considering the energetic aspects of the
front motion is chosen [wiegold2019]. The front movement, represented by the kinetic energy Ekin =
2 k T ξeq v+ is caused by the difference in the energy in front of and behind the front, each consisting of
1 2
several contributions denoted by superscript "−" behind the front and superscript "+" in front of the
front. Thus,
Eb− + Ee− + Eκ− − Ee+ + Ev+ = Ekin . (3)
Behind the front, Eb− = −k T Cb ξeq denotes a reduction in the free energy due to the binding of receptors,
with k being theBoltzmann
constant, T the absolute temperature and Cb a reduction factor. The term
ξeq
Ee = k T ξeq ln ξ0 represents the energy associated with the entropy of receptors. Here, the initial
−
binder density ξ0 is adopted as a reference value. The term Eκ− = 12 k T B κ2 represents the energy related
with bending, where κ is the curvature
of the membrane and where B is the bending stiffness. In front
ξ+
of the front Ee = k T ξ+ ln ξ0 represents the energy associated with the entropy of receptors. The
+
term Ev+ = 12 k T ξ+ vr2 represents the kinetic energy of the receptors in front of the front moving towards
the adhesion area, with vr being their corresponding velocity. Finally, the second boundary condition is
expressed in terms of the receptor density ξ(x, t) and the front velocity v+ by

ξeq 1 ξ+ 1 1
−ξeq Cb + ξeq ln + Bκ2 − ξ+ ln + k T ξ+ vr2 = 2
ξeq v+ . (4)
ξ0 2 ξ0 2 2
The corresponding non-linear system of equations is implemented by using the finite difference method
and is solved numerically by a Newton Raphson scheme.
2 Cooperativity
As soon as receptors of two opposing surfaces create bonds they will smooth out the membranes such
that they become more conform to each other [weikl2009]. This makes it easier for additional receptors
to create bonds and strengthens the adhesion between the two surfaces. This effect during the edocytosis
process is called cooperativity. The cooperativity reduces the total amount of required receptor bonds in
order to connect the virus with the cell according to
κ 2 2
ξeq_req = c( )l K [ξeq ]2 [ξ0 ]2 . (5)
kB T we pl
Here, ξeq_req is the required amount of receptors that need to bind, c is a dimensionless prefactor, lwe
is the interaction range between two opposing receptors and Kpl is the two dimensional equilibrium
constant.
3 Numerical results
The chosen numerical example models the uptake of a spherical virus. The evolution of the receptor
profile of the cell as well as the motion of the adhesion front are shown in Fig. 1. At the beginning of the
process the receptor density at the front experiences a strong decrease and reaches ≈ 50% of its initial
value after 300 time steps. This decrease weakens as the time progresses. During the whole process the
adhesion front advances continuously, while its velocity decreases towards a limit value. The numerical
402
results are strongly influenced by the mobility parameter m which measures how fast the cell can provide
receptors to the adhesion zone. Further important factors are the initial receptor density of the virus as
well as of the cell [wiegold2019].
Figure 1. Receptor density over the cell surface for the time steps 0 - 1500. Chosen process parameters are
m = 1µm2 /s, Cb = 5, B = 30, ξeq = 4800 1/µm2 , ξ+ = 1000 1/µm2 , virus diameter D = 0.05µm.
Cooperativity is incorporated into the simulations by considering a patch of receptors with a distinct
interaction range surrounded by receptors with a different interaction range. The two considered in-
teraction ranges are lwe = 1.2 nm in the inner patch, which leads to a required receptor density of
ξeq_req = 3262 1/µm2 and lwe = 1 nm for the surrounding receptors, which leads to a required receptor
density of ξeq_req = 2265 1/µm2 . The results of the simulations for the described setup are shown in
Fig. 2. The top row shows the density profile over the cell membrane for different time steps up to the
state when the whole virus entered the cell. The velocity of the front, the position of the front and the
receptor density at the front are presented in the bottom row of Fig. 2.
Figure 2. The required density is calculated taking into account the cooperativity. Chosen process parameters are
the same as in Fig. 1. Additionally ξeq_req = 3262 1/µm2 in the inner patch and ξeq_req = 2265 1/µm2
in the surrounding area. The upper row shows the receptor density over the cell surface for different
time steps. The bottom row shows the velocity of the front, position of the front and receptor density
at the front over time.
403
The different patches are clearly differentiated in the diagrams by either a jump or a kink in the corre-
sponding diagrams. The velocity is most affected by the change of the required density. In the area with
a smaller interaction range less receptors are required, which significantly increases the velocity of the
process.
4 Conclusions
The present study focuses on the simulation of the influence of the binder mobility in the viral entry
driven by receptor diffusion. The process is described by a diffusion differential equation accompanied
by two boundary conditions. The two boundary conditions describe the flux balance and the energy
balance at the adhesion front. The proposed moving boundary problem enables an efficient numerical
simulation of the process and provides the basis for the consideration of additional relevant aspects like
cooperativity. Bending of the cell in front of the adhesion front and an alternative expression for bending
lipid bilayers are envisaged as topics of the future work.
Acknowledgments
We gratefully acknowledge the financial support of the German Research Foundation (DFG), research
grant No. 385960030, and the Austrian Science Fund (FWF), research grant No. I 3431-N32.
404

Optimization-based comparison of different anisotropic

models for growth and remodeling in arterial walls
∗
Anna Zahn and Daniel Balzani
Chair of Continuum Mechanics, Department of Civil and Environmental Engineering,
Ruhr-Universität Bochum, Universitätsstraße 150, 44801 Bochum, Germany
Abstract. The simulation of growth and remodeling processes is an elegant way to obtain approximations
for residual stresses and fiber orientations in arterial walls. In order to enable a comparison of different
model variants within a framework of combined growth and remodeling, and to assess the assumptions on the
underlying fundamental mechanisms, a method based on the optimization of model parameters is developed.
The minimization of a mechanobiologically motivated objective function permits to evaluate the approaches
with respect to their ability of effectively reducing stress peaks and stress inhomogeneities in the arterial wall.
Introduction
Unfortunately, residual stresses and fiber orientations in arterial walls can currently hardly be measured
in vivo. However, they considerably affect the mechanical behavior of arterial tissues and have thus to be
known in order to perform reliable patient-specific numerical simulations. By modeling the adaptation
processes of arteries to their mechanobiological environment, an approximation of residual stresses and
fiber orientations can be obtained automatically as a result of arterial growth and fiber reorientation. The
combined framework for stress-driven growth and remodeling presented in [8] provides a set of variable
formulations, which are based on different hypotheses on the mechanobiological mechanism of stress
reduction. In order to identify the most probable of these mechanisms, a quantitative comparison of the
model variants has to be enabled. To this aim, the growth parameters are computed by minimizing an
objective function, which is formulated such that its minimum reflects the energetically most efficient
reduction of stress peaks and inhomogeneities in the arterial wall. In Section 1, a summary of the general
framework is given. Section 2 focuses on the optimization-based comparison of model variants within
this framework and gives numerical results. The contribution ends with a short conclusion.
1 General growth and remodeling framework
Following Rodriguez et al. [7], the deformation gradient F = Fe Fg is multiplicately decomposed into a
growth part Fg , which is associated to a stress-free volume change by factor Jg = det[Fg ], and an elastic
part Fe . With regard to the orthotropic behavior of soft biological tissues, the growth tensor itself is
(3) (2) (1)
decomposed according to Fg = Fg Fg Fg into three parts, which are related to a set of perpendicular
(a)
directions Ag (a = 1, 2, 3) and to the internal variables ϑ(a) , which determine the amount of growth
associated to these directions. Based on the assumption that adaptation processes in soft biological tissues
are stress-driven [3], the elastic part of the Mandel stress tensor Ce Se =·· Σe with Ce = Fg−T CFg−1 and
Se = 2 ∂ψ/∂Ce is chosen as driving quantity for growth and remodeling as suggested by Himpel et al. [5].
(a)
The directions Ag are then computed as principal directions of Σe and growth related to those directions
is assumed to reduce the associated principal stresses. The growth factors ϑ(a) are defined by a coupled
(a)
set of evolution equations ϑ̇(a) = kϑ (ϑ(a) ) φ(a) (Σe ), involving a growth driving quantity φ(a) formulated
(a)
in terms of Σe and a growth function kϑ taken from Lubarda & Hoger [6]. This function assures that
the growth factors ϑ are within a range of [ϑ−
(a) + − +
(a) , ϑ(a) ] with ϑ(a) < 1 and ϑ(a) > 1.
∗ Corresponding author: daniel.balzani@ruhr-uni-bochum.de
405
A polyconvex material model for fiber-reinforced materials is used to describe the basic material behav-
ior [1]. Each of the two included fiber families is assumed to continously reorient in response to local
demands. The target fiber orientation vectors are defined based on the principal directions and eigenval-
ues of the stress tensor Σe in analogy to the approach of Hariton et al. [4]. The reorientation is regulated
by the angles η (1) and η (2) between the existing and the target fiber vectors, which are supposed to evolve
towards zero over time as described by the evolution equation η̇ (a) = kη (η (a) ). The remodeling velocity
is controlled by the parameters of the remodeling function kη (η (a) ) = −kη+ ln(m+ η |η
(a)
| + 1), cf. [8].
Based on hypotheses on the mechanical mechanism of stress reduction, different types of the multiplicative
parts of the growth tensor and the associated driving forces are imaginable. Here, the growth tensors
Fg(a) = ϑ(a) I for isotropic growth, (1a)

Fg(a) = ϑ(a) I + 1 − ϑ(a) A(a)
g ⊗ Ag
(a)
for growth perpendicular to A(a)
g and (1b)

Fg(a) = I + ϑ(a) − 1 A(a)g ⊗ Ag
(a)
for growth in the direction of A(a)
g (1c)
are considered. Whereas isotropic growth is supposed to be least realistic, an increase of the cross section
as well as growth in the direction of the stress might be adequate. A set of driving forces is given by
φ(a) (Σe ) = Σe : I, (2a)

φ(a) (Σe ) = Σe : Mg(a) with Mg(•) = Ag(•) ⊗ A(•)
g , (2b)
(a)
φ (Σe ) = hΣe : Mg(a) i and (2c)
(
(I) (II)
1
Σ e : M g + Mg for ΣIe > 0, ΣII
e >0
φ(a) (Σe ) = 2 (I)
, (2d)
Σe : M g for ΣIe > 0, ΣII
e ≤0
where the indices I/II point out that the 1st /2nd principal directions of Σe are referred to, and ΣIe /ΣII
e are
the eigenvalues of Σe . Besides the isotropic stress measure in Eq. (2a), a projection of Σe in the direction
of Ag is considered in Eq. (2b). The case that only tensile values of these stresses provoke growth is
reflected in Eq. (2c). Eq. (2d) represents the average stress state within the plane spanned by the fibers.
2 Optimization-based comparison of model variants
With the aim of finding combinations of model components which are suited best to improve the load-
bearing behavior, a quantitative comparison of model variants has to be enabled. To this end, the
growth and remodeling parameters have to be set specifically for each model variant, such that the
best possible particular behavior can be achieved. This is done by minimizing a biologically motivated
objective function, which then also serves as comparative value between the model variants. Most of the
parameters mainly affect the temporal behavior but rarely the state of steady growth. The optimization
is therefore restricted to the limiting values ϑ+
(a) for positive growth – all other parameters are throughout
set as specified in Tab. 1. As stated in [8], the ratio of internal pressure and axial strain is an additional
important factor affecting the resulting stresses. To accommodate this, the level of the axial displacement
is either considered as optimization variable as well or an averaging over different levels is performed.
2.1 Objective function
Provided that
P a steady growth state under constant load is reached, the value of the objective func-
tion fobj = ωi qi serves as comparative measure between different model variants. The arguments
max |σϕ | max |σz | mean |σϕ − σz |

r r r
q1 = , q2 = , q3 =
σ̃peak σ̃peak mean σϕ ,
r

max |σϕ | − mean σϕ max |σz | − mean σz
q4 = r r
, q5 = r r
, q6 = mean Jg − 1
σ̃diff σ̃diff r
denote optimization objectives which are claimed to be minimal in a state of steady growth. The operator
“mean” denotes the volume average of a quantity over the radial direction r, and the stresses σ̃peak =
406
100 kPa and σ̃diff = 50 kPa are introduced as normalizing values. All expressions are normalized such
that they are unitless and of same order of magnitude. Their mechanobiological motivation is, that the
aim of efficient adaptation processes is a reduction of stress peaks (q1 , q2 ), differences (q3 ) and gradients
(q4 , q5 ) at a low level of energetic costs, i. e. a low volume change induced by growth (q6 ). Assuming that
all objectives are equally important, all weighting factors ωi are set to ωi = 1.
2.2 Exemplary results
A series of model variants assuming the same combination of growth tensor and driving force for each of the
(a)
three directions Ag is analyzed and compared to isotropic growth. Exploiting the rotational symmetry
of idealized arterial segments, all simulations are reduced to the radial dimension using rotationally
symmetric 1D finite elements. The initial fiber angles are set to ± 30◦ and the internal pressure to
p = 120 mmHg. A prescribed axial displacement uz is either chosen among 0 %, 10 % or 20 % of the
initial length or determined by optimization. All other, unvaried parameters are listed in Tab. 1.
Table 1. Parameters of the material model [1], adjusted to human media in [2], parameters of the remodeling
(a)
function kη and unvaried parameters of the growth function kϑ , set equally for all directions.
c1 1 2 α1 α2 kη+ m+
η kϑ+ m+
ϑ ϑ− kϑ− m−
ϑ
17.5 kPa 499.8 kPa 2.4 30 001.9 kPa 5.1 0.6 s−1 5.0 1.0 s−1 3.0 0.95 1.0 s−1 3.0
Table 2. Type of growth tensor Fg

(a)
Table 3. Type of driving force φ(a) (Σe )
1 isotropic growth Eq. (1a) 2 isotropic Eq. (2a)
(a)
2 growth perpendicular to Ag
(a)
Eq. (1b) 4 stress in the direction of Ag Eq. (2b)
4
(a)
growth in direction of Ag Eq. (1c) 5 stress within the plane of the fibers Eq. (2d)
(a)
6 tensile stress in the direction of Ag Eq. (2c)
ref ref
For the relative comparison of the different approaches, a performance measure P = −(fobj − fobj )/fobj
is defined, which represents the improvement of each approach with respect to the reference case where
ref
no growth and remodeling are included. This reference value fobj depends on the height of the axial
displacement, see Fig. 1d. Positive values of P are associated with model variants which indeed lead to
improvements in the mechanobiological state: the higher P , the stronger are the improvements.
The optimization results are given in Fig. 1a–c. The denotations on the horizontal axis consist of up
to three pairs of numbers for each approach, which are associated to the type of growth tensor and
driving force for the three directions. Tab. 2–3 explain those numbers and refer to the correspondent
equations. In Fig. 1a, the performance measure obtained when optimizing growth parameters ϑ+ (a) and
axial loading uz is illustrated. This optimization scenario reveals the best possible performance of the
models. All considered approaches lead to an improvement compared to the reference case, and most
anisotropic approaches perform better than isotropic growth (denoted by “12”). However, the axial
load might vary naturally within a physiological range, such that Fig. 1a alone is not sufficient for an
assessment. If a single set of growth parameters is identified for an averaged value of the objective
function using uz ∈ [0 %, 10 %, 20 %], the results shown in Fig. 1b are obtained. They represent the mean
model performance and are in most cases significantly lower than in Fig. 1a. In Fig. 1c, the sensibility of
the approaches with respect to the axial displacement is shown. Here, the parameters are optimized for
each load level separately, which explains the higher performance measures compared to Fig. 1b.
Conclusion
A method for the comparison of different approaches within an existing framework of combined growth
and fiber reorientation in arterial walls has been proposed. It relies on the evaluation of a mechanobiolog-
ically motivated objective function at steady growth state and should in principle enable a comparative
assessment of the fundamental assumptions on which the individual model variants are founded.
With regard to the numerical results presented here, it is obviously not possible to draw a clear conclusion
concerning the most probable growth mechanism so far. High performance measures for the average
407
0.6 0.6
optimization parameters: .
0.5 0.5
ϑ+(a)
averaged over uz,i
0.4 0.4
P in –
P in –
0.3 0.3
0.2 0.2
optimization parameters: ..
0.1 u z , ϑ+ 0.1
(a)
0.0 0.0
12
141414
222222
242424
424242
444444
12
141414
222222
242424
424242
444444
161616
252525
262626
454545
464646
161616
252525
262626
454545
464646
a growth model b growth model
0.6 6
uz = 0 %
0.5 uz = 10 % 5
0.4 uz = 20 % 4
f ref in –
P in –
0.3 3
obj
0.2 2
0.1 1
0.0 0
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0
12
141414
222222
242424
424242
444444
161616
252525
262626
454545
464646
axial displacement in %
c growth model d
Figure 1. Performance measure P for optimization of a subset of different approaches with respect to a) the best
model performance, b) the average model performance and c) the best model performance for a given
axial loading. Subfigure d shows the reference values of the objective function.
model performance are obtained for non-isotropic growth tensors in combination with all driving forces
except the one given in Eq. (2b). To get more insight into the results, the robustness of the individual
approaches is worth to be considered in further examinations. If this is still not sufficient to obtain a
clearer statement, it might on the one hand be necessary to vary the weighting factors for the arguments
of the objective function or to take further important effects like smooth muscle activity into account. On
the other hand, it might be a hint on faulty model assumptions, for example the one that all adaptation
processes are stress-driven.
Acknowledgments
The authors thank the DFG (German Science Foundation) for funding in the project BA 2823/13-2.
References
[1] D. Balzani, P. Neff, J. Schröder, and G. A. Holzapfel. Int. J. Solids Struct. 43 (2006), pp. 6052–6070.
[2] D. Brands, A. Klawonn, O. Rheinbach, and J. Schröder. Comput. Methods Biomech. Biomed. Engin.
11 (2008), pp. 569–583.
[3] E. Comellas, A. Carriero, M. Giorgi, A. Pereira, and S. J. Shefelbine. Modeling the Influence of
Mechanics on Biological Growth. In: Numerical Methods and Advanced Simulation in Biomechanics
and Biological Processes. Ed. by M. Cerrolaza, S. J. Shefelbine, and D. Garzón-Alvarado. Elsevier,
2018. Chap. 2, pp. 17–35.
[4] I. Hariton, G. deBotton, T. C. Gasser, and G. A. Holzapfel. Biomech. Model. Mechan. 6 (2007),
pp. 163–175.
[5] G. Himpel, E. Kuhl, A. Menzel, and P. Steinmann. Comput. Model. Eng. Sci. 8 (2005), pp. 119–134.
[6] V. A. Lubarda and A. Hoger. Int. J. Solids Struct. 39 (2002), pp. 4627–4664.
[7] E. K. Rodriguez, A. Hoger, and A. D. McCulloch. J. Biomech. 27 (1994), pp. 455–467.
[8] A. Zahn and D. Balzani. Z. Angew. Math. Mech. 98 (2018), pp. 2072–2100.
408
MS 21: Representative Volume Elements: Microstructure Analysis and Homogenization

Application of Voronoi tessellations to model cellular

aerogels
1
Rajesh Chandrasekaran ∗
, Markus Hillgärtner1 , Barbara Milow2 , Ameya Rege2 and Mikhail Itskov1
1 Department of Continuum Mechanics, RWTH Aachen University, Aachen, Germany
2 Institute of Materials Research, German Aerospace Center, Cologne, Germany
Abstract. Many aerogels are characterized by a fibrillar morphology. Recently proposed constitutive models
have shown good predictive capabilities in describing the mechanical behavior of such aerogels. On the one
hand, these models account for the cellular nature and adhere to the cell-size distributions, while on the
other hand they are based on the assumption of the network made up of idealized square-shaped cells. The
heterogeneous cellular morphology of aerogels can computationally be described by a Voronoi tessellation-
based approach. Accordingly, the random cell shapes within the network are accounted for by means of the
pore-size distribution data obtained from experiments. To this end, 2-d and 3-d periodic representative volume
elements (RVEs) are generated from the cellular structure obtained using Voronoi tessellations and simulated
under compression using finite element analysis.
Introduction
Aerogels exhibit unique physical properties such as high porosity, high specific surface area, low thermal
conductivity and low density. Aerogels are synthesized by sol-gel processing from molecular precursors [3].
Amongst many kinds of aerogels, polysaccharide-based ones are characterized by a fibrillar morphology
with the cellular-like microstructure. The microstructure of biopolymers can be tailored by means of
different material sources, extracting and processing methods, and also by synthesis and drying routes [10].
For example, the density and pore size distribution of biopolymer aerogels are observed to depend on
the biopolymer concentration. Accordingly, the mechanical properties can be directly evaluated based on
synthesis parameters.
The mechanical and structural properties of biopolymer aerogels have recently been studied. Rege et
al. [7] developed a micromechanically motivated constitutive model to describe the mechanical behavior
of biopolymer aerogels. The model formulations were based on the assumption that the aerogel network
consists of continuously distributed idealized square shaped microcells. Recently, the model was further
improved and generalized to describe the mechanical behavior and properties of other biopolymer aerogels,
like pectin and k-carrageenan [8]. Although this constitutive model demonstrated good agreement with
experimental data, the assumption of the square shaped microcells remains an idealization. The shape
of microcells throughout the aerogel network is highly randomized and not restricted to quadrangle. In
order to create a more realistic geometry, a Voronoi tesselation-based approach can be applied. One of
the advantage of this approach is that, it not only accounts for the randomized cell size distribution, but
also for the randomised cell shapes.
In this contribution, a recently reported model [9] characterizing biopolymer aerogel morphologies by
means of a Voronoi tessellation-based algorithm is presented and extended. The Voronoi geometry is
designed to fit the obtained pore size distribution data from experimental nitrogen sorption tests. The
homogenized macroscopic response of the model was simulated using the concept of representative volume
element (RVE) employed in combination with periodic boundary conditions (PBCs). The 2-d model is
extended to 3-d in order to obtain realistic microstructures.
∗ Corresponding author: chandrasekaran@km.rwth-aachen.de
409
1 Methods
1.1 Voronoi tessellation of aerogels
A Voronoi tessellation technique decomposes the entire domain in m-dimensional space into n distinct
regions with random set of points [1]. Each region Ri (i = 1, 2, ..., n) forms a Voronoi cell containing
exactly one generating point, known as seed Si (i = 1, 2, ..., n). For each seed Si there exist a region
consisting of all points closer to that particular seed Si than to any other seed Sj (j = 1, 2, ..., n, i 6= j).
Based on the Euclidean metric (|| • ||2 ), the i-th Voronoi cell generated from n seeds can be expressed as

m

Ri = x ∈ R ||x − Si ||2 ≤ ||x − Sj ||2 ∀ j = 1, 2, ..., n : j 6= i, (1)
where x represents the position of any point in m-dimensional space. In 2-d space (m = 2), the tessellation
results in n Voronoi cells in the shape of convex polygon. Each vertex of i-th cell is equidistant from
two nearest seeds. In the present approach, the Voronoi structure was generated in MATLAB [6], which
utilizes the open source software Qhull [2] to create 2-d Voronoi tessellations. In order to create a periodic
cellular microstructure, the Voronoi seeds were repeated in a 3 × 3 pattern. To obtain a geometry
resembling microstructure of aerogels, the Voronoi seeds were chosen in accordance with the measured
cell size distributions resulting from the experimental data (available in [9]).
1.2 Finite element modelling
From the Voronoi structures generated in MATLAB for three different cell size distribution of aerogels,
square shaped RVEs of the dimension 250nm x 250nm were created. Each RVE has a fully periodic
cellular pattern in order to apply PBCs. Furthermore, the edges of the Voronoi cells are represented as
FE beam elements with a cross section reproducing the cell wall fibres in the biopolymer aerogel network.
To this end, the edges of the polygonal cells were imported into LS-DYNA as beam elements with circular
cross section. The Belytschko beam element formulation was used [4]. A linear elastic material model
(*MAT_ELASTIC) with Young’s modulus of 12 GPa was defined to each of the beam elements. A global
contact using *CONTACT_AUTOMATIC_GENERAL with 0.1 friction coefficient was defined to the
entire RVE. This contact formulation is suitable for beam-to-beam contacts [5]. The Voronoi geometry
constituting of LS-DYNA beam elements representing the microstructure of the aerogel with a cellulose
concentration of 5 wt.% is shown in Fig. 1a.
(a) RVE in the reference state (b) Distribution of the Axial force (in nN) after com-
pression up to 25% strain.
Figure 1. A 2-d aerogel RVE in LS-DYNA.
410
2 Results
Voronoi tessellation-based RVEs for different cell-size distributions and cell wall fiber diameters were
generated and imported as beams in LS-DYNA. The dummy node (which controls the periodic boundary
nodes) was subjected to a compression of 25% and the corresponding deformation of the RVE was studied.
The contour plot of the axial force resulting in cell wall fiber is shown in Fig. 1b for aerogel with 5 wt.%
cellulose concentration and a diameter of 6.4 nm .
With the increasing concentration of cellulose in aerogels the fiber diameter of the cell wall increases.
Aerogels with 3, 5 and 7 wt.% cellulose concentration have diameter 5.5, 6.4 and 7.6 nm, respectively.
By increasing the fiber diameter, the volume of fiber solid content increases, thereby reducing the specific
surface area and affecting the pore size distribution [8]. Accordingly, the cell wall fiber diameter and
the pore space can be controlled during synthesis process. By this means, tailoring the target properties
and response of cellulose aerogels (CAs) using computational results is of special interest. Therefore, the
sensitivity of the bulk material properties to these structural parameters was investigated and presented
in Rege et al. [9].
Figure 2. Model prediction of 2-d RVE-based analysis
The Voronoi based RVE model for different CA concentration with corresponding PDFs and fiber diameter
was subjected to a compression of 25%, in order to investigate their actual bulk behavior. The force-
strain curve predicted by the model for 3, 5 and 7 wt.% CAs are plotted in Fig.2. Accordingly, the 2-d
periodic model effectively captures the trend of force-strain curves obtained in experiments. A region
of linear elastic response is observed at low strains due to local bending and buckling of the cell wall
fibers. Subsequently, yielding and damage occurs. Furthermore, the contact between the FE beams
(i.e. cell wall fibers) affects the force-strain curve. The contact strain (i.e. the strain at which the cell
wall contacts with each other within the microstructure) occurs between 12% and 18% depending on the
different cellulose concentration.
3 Discussion
In this first approach, only a qualitative agreement is obtained. This is primarily because the out–of–plane
connectivity is not accounted for due to the 2-d periodic RVE-based model. Although, the biopolymer
aerogel network consists of open cellular structures, it has non-trivial out–of–plane connectivity which
requires 3-d representation. The 2-d approach is incapable of capturing the bulk material response of a
physical 3-d aerogel specimen in comparison to the macroscopic experimental data. Accordingly, a 3-d
periodic RVE based on Voronoi structure was generated and simulated in LS-DYNA using larger RVE
size, as a headway to this approach (Fig. 3). A Voronoi cell in 3-d space is a convex polyhedron formed by
polygonal faces. Note that the cross section of a 3-d Voronoi tessellation is not based on the 2-d Voronoi
tessellation. Therefore, the cell size distribution data used for generation of 2-d Voronoi does not account
for the 3-d Voronoi structure.
411
(b) Deformed state

(a) Reference state
Figure 3. 3-d voronoi based RVE in LS-DYNA
4 Conclusion
This paper describes the modeling of 2-d Voronoi tessellation-based RVEs for biopolymer aerogels and
its respective finite element simulation. Application of periodic boundary conditions to the RVEs are
helpful in approximating the bulk properties. The force-strain response under compression of periodic
2-d model is qualitatively in good agreement with the experimental results of biopolymer aerogels [8].
As an extension to this approach, 3-d Voronoi tessellation-based RVEs was created and tested. By using
this approach, quantitative agreement, along with a qualitative one, could be obtained. Such 3-d RVEs
will be generated, adhering to the experimental pore-size distribution data, and an analysis of the bulk
properties based on their performance will be presented elsewhere.
References
[1] F. Aurenhammer. Voronoi diagrams—a survey of a fundamental geometric data structure. ACM
Computing Surveys (CSUR) 23 (1991), pp. 345–405.
[2] C. B. Barber, D. P. Dobkin, and H. Huhdanpaa. The quickhull algorithm for convex hulls. ACM
Transactions on Mathematical Software (TOMS) 22 (1996), pp. 469–483.
[3] J. L. Gurav, I. K. Jung, H. H. Park, E. S. Kang, and D. Y. Nadargi. Silica aerogel: synthesis and
applications. Journal of Nanomaterials 2010 (2010), p. 23.
[4] J. O. Hallquist. LS-DYNA theory manual. Livermore software Technology corporation 3 (2006),
p. 6.
[5] LS-DYNA Keyword User’s Manual and I Volume. Version 11. Livermore Software Technology Cor-
poration (2018), p. 11.
[6] MATLAB. version R2018b. M. Natick: The MathWorks Inc., 2018.
[7] A. Rege, M. Schestakow, I. Karadagli, L. Ratke, and M. Itskov. Micro-mechanical modelling of
cellulose aerogels from molten salt hydrates. Soft Matter 12 (2016), pp. 7079–7088.
[8] A. Rege, I. Preibisch, M. Schestakow, K. Ganesan, P. Gurikov, B. Milow, I. Smirnova, and M. Itskov.
Correlating Synthesis Parameters to Morphological Entities: Predictive Modeling of Biopolymer
Aerogels. Materials 11 (2018), p. 1670.
[9] A. Rege, M. Hillgärtner, and M. Itskov. Mechanics of biopolymer aerogels based on microstructures
generated from 2-d Voronoi tessellations. The Journal of Supercritical Fluids 151 (2019), pp. 24–29.
[10] S. Zhao, W. J. Malfait, N. Guerrero-Alburquerque, M. M. Koebel, and G. Nyström. Biopolymer
aerogels and foams: Chemistry, properties, and applications. Angewandte Chemie International
Edition 57 (2018), pp. 7580–7608.
412

Microstructural convergence analysis as basis for an ef-

cient and accurate two-scale FE-FFT-based simula-
tion approach
∗ 1 1 1 2,3 1
Christian Gierden , Julian Kochmann , Johanna Waimann , Bob Svendsen and Stefanie Reese
1 Institute of Applied Mechanics, RWTH Aachen University, Aachen, Germany

2 Material Mechanics, RWTH Aachen University, Aachen, Germany
3 Microstructure Physics and Alloy Design, Max-Planck-Institut für Eisenforschung GmbH, Düsseldorf, Germany
Abstract. We present an ecient and accurate solution scheme for a FE-FFT-based two-scale simulation
which is based on a microstructural convergence analysis leading to a minimal number of grid points, which
needs to be considered in terms of the two-scale simulation. While performing the entire simulation with this
coarse discretized microstructure, the macroscopic deformation gradient is stored at every point of particular
interest. Finally, in a post-processing step, this macroscopic deformation gradient is applied to a ne discretized
microstructure to generate highly resolved microstructural data.
Introduction
To capture all the individual microstructural eects of a complex and heterogeneous material behavior in
a structural nite element simulation, the usage of a two-scale simulation approach is necessary. In this
context we use the FE-FFT-based method (e.g. [3]). Prakash and Lebensohn already showed that this
new method is more ecient than the common FE2 method ([10]). Nevertheless, highly resolved two-scale
simulations of complex structures are still barely possible due to the high computational eort of both
methods. One possibility to overcome this high computational eort is to combine the FFT-solvers with
model order reduction techniques in Fourier space (e.g. [5]). Instead of that, we present the versatile
use of an ecient FE-FFT-based simulation approach based on a microstructural convergence analysis in
real space [4], which is also presented in a more detailed way in [1]. As a numerical example we present
a bending test of a cantilever beam with polycrystalline material behaviour.
1 Two-scale simulation
Considering nite strains, the deformation of a macroscopic body B̄ is described by the macroscopic
deformation gradient F̄ (X̄) which is related to position X̄ ∈ B̄ in the reference conguration. The
macroscopic material response, e.g. the macroscopic rst Piola-Kirchho stress P̄ (X̄), at these points
is given by the volume averaged response of a microstructure simulation. To get these microstructural
responses, the microscopic boundary value problem DivP (F (X̄, X)) = 0 (neglecting body forces) needs
to be solved. Here, X is the local coordinate corresponding to the microstructure B and P (F (X̄, X)) is
the microscopic rst Piola-Kirchho stress depending on the total deformation gradient
F (X̄, X) = F̄ (X̄) + H̃(X) (1)
which is additively split into F̄ (X̄) and a microscopic spatially uctuating part H̃(X) = ∇ũ(X). The
unknowns ũ(X) of the microscopic boundary value problem represent the uctuating displacement eld.
Due to a better readability, the dependencies on X̄ and X are neglected in the following.
∗ Corresponding author: christian.gierden@rwth-aachen.de
413
2 FFT-based microstructure simulation
To overcome the dependence of the stresses on spatially varying material properties the polarisation stress
τ = P (F ) − C0 : F (2)
is introduced which is the dierence between the stresses in the inhomogeneous microstructure and the
stresses in a homogeneous reference material C0 . This polarisation stress leads to the reformulated
microscopic balance equation Div(C0 : F ) = −Div(τ ). To solve this balance equation Moulinec and
Suquet [7] proposed to use Green's function which then leads to the Lippmann-Schwinger equation to
calculate the total deformation gradient
L
F = F̄ − 0 ∗ τ (3)
where ∗ denotes the convolution integral. To solve the Lippmann-Schwinger equation it is transformed
into Fourier space where it reads
(
L̂0
F̂ (ξ) = − (ξ) τ̂(ξ) for ξ 6= 0 (4)
F̄ for ξ = 0
and the Green operators Γ̂0ijkl (ξ) = Ĝ0ik (ξ) ξj ξl and Ĝ0ik (ξ) = (C0ijkl ξj ξl )−1 are explicitly known and just
depending on the wave vectors ξ and the homogeneous reference material C0 . To solve this problem, we
use the ecient and robust FFT- and conjugate gradient-based solution scheme which is introduced in
[2] in the context of nite strains and based on the linearized Lippmann-Schwinger equation
h L i L
I + 0 ∗ C? (i) − C(0) : ∆F (i+1) = F̄ − F (i) − 0 ∗ τ (i) (5)
with the local tangent operator for the i-th iteration step C? (i) = ∂F P |F (i) .
3 Viscoplastic polycrystalline material formulation
In a nite strain setting we assume a multiplicative split of the total deformation gradient F = Fe Fp
into an elastic part Fe and a plastic part Fp . Since elastic deformations in polycrystals are generally
small, a Saint Venant-Kirchho model is used to capture these elastic deformations which leads to the
second Piola-Kirchho stress Se = 12 C : (Ce − I) in the intermediate congurraten with the left Cauchy-
Green tensor Ce = FeT Fe and the elasticity tensor C. Considering polycrystalline materials with nslip
slip systems, plastic deformations will evolve if the resolved shear stress τα (Me ) = Me : (dα ⊗ nα ) in
any slip system α exeeds the critical shear stress τ C (γacc ). Here, Me = Ce Se is the mandel stress which
is related to the plastic velocity gradient Lp = F˙p Fp−1 . The vectors dα and nα are the slip direction and
the plane normal unit vectors, respectively. Following Peirce et al. [8], the plastic velocity gradient
nslip
X
Lp = γ̇ dα ⊗ nα (6)
α=1
is dened as the superposition of the plastic slip rates γ̇α in each slip system. The evolution of γ̇α is given
by the Perzyna-type ow rule
p
|τα | − τ C (γacc )
γ̇α = γ̇0 sgn(τα ) (7)
τD
with the reference shear rate γ̇0 , the drag stress τ D and the rate sensitivity parameter p [9]. The
critical P
resolved
R shear stress τ C (γacc ) = τ0C + qh (γacc ) is depending on the accumulated plastic slip
γacc = α |γ̇α |dt via the isotropic hardening function qh (γacc ) which is dened to be a Voce-type
hardening behaviour:
qh (γacc ) = τ0c + (τ∞ − τ0 ) tanh[((h0 − h∞ ) γacc )/(τ∞ − τ0c )] + h∞ γacc (8)
The ow rule (7) is discretized by using an implicit time integration scheme (backward Euler) with
continuously dierentiable transition between the powerlaw and a linear approximation (see [11]).
414
4 Solution strategy and a numerical example
Let us consider a bending test of a cantilever beam consisting of pure copper with the slip systems
dα = {111} and nα = {110}. The material parameters for pure copper are given in table 1.
C11 [GPa] C12 [GPa] C44 [GPa] τ0c [MPa] τ D [MPa] τ∞ [MPa] h0 [MPa] p [-]
170.2 114.9 61.0 0.1 10 31.0 50.0 20
Table 1. Used material parameters for Copper [6] with γ̇0 = 10−3 1/s.
In terms of a pre-processing step we consider a microstructure with 100 grains and a random texture. This
microstructure is submitted to three dierent deformation gradients (F ◦ , F 4 , F ) with Fmax = 0.15
and ∆t = 5 · 10−8 s. To determine the minimal discretization Nmin of this microstructure, six dierent
discretizations with N = 7, 15, 31, 63, 127 and 255 grid points in each direction are considered. The result
is given in gure 1. It can be seen, that the error Ē in the norm of the macroscopic rst Piola-Kirchho
stress P̄ of any discretization compared to the converged solution P̄ with N = 255 grid points in each
direction is always smaller than 1% for a minimal discretization of Nmin = 15 grid points in each direc-
tion. For a convergence analysis concerning dierent textures see [1].
3
1 Fmax
F◦ =
0 1
Ē[%] = ||P̄ ||L2 /||P̄||L2 − 1
2
1 + Fmax 0
F4 = 4
1
0 1/F11

1 + Fmax Fmax
F =
0 0 1/F11
-1
[111]
-2
Texture [001] [011]
-3
15 63 127 255
number of grid points in each direction

Figure 1. Convergence analysis in terms of the norm of the macroscopic rst Piola-Kirchho stress P̄ .
Concerning the two-scale simulation it is then possible to get suciently accurate macroscopic results
by considering only those 15 grid points in each direction. In gure 2 the boundary conditions and the
equivalant stress σ̄eq of the macroscopic boundary value problem is shown.
0 σ̄eq [MPa] 1.5
lx = 10 mm
ly = 1 mm
u = 5 mm
u̇ = 5 mm/s
0 γacc [−] 0.4
Figure 2. Macroscopic equivalant stress σ̄eq computed with Nmin = 15 and highly resolved microscopic accumu-
lated plastic strain γacc computed with N = 255 grid points in each direction.
415
During the simulation the macroscopic deformation gradient of the lower right corner is stored. In a post-
processing step it is then used to compute highly resolved microstructural results, e.g. for the microscopic
accumulated plastic slip, with N = 255 grid points in each direction (see gure 2).
5 Conclusions
The proposed ecient and accurate two-scale solution scheme is decomposed into a pre-processing, a
processing and a post-processing step. After nding the minimal number of grid points, which is needed
to get suciently accurate macroscopic quantities, the entire two-scale simulation is performed with
this relatively coarse discretization. Finally, in a post-processing step, high-delity solutions may be
generated for any integration point of particular interest. By using this solution scheme high-dility
two-scale simulations can be performed in a reasonable time period.
Acknowledgements
The work is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation)
Projektnummer 223500200 TRR 136. We greatfully acknowledge the nancial support of the subpro-
jects M03 and M05.
References
[1] C. Gierden, J. Kochmann, J. Waimann, B. Svendsen, and S. Reese. Ecient two-scale FE-FFT-
based mechanical process simulation of elasto-viscoplastic polycrystals at nite strains. In prepara-
tion (2019).
[2] M. Kabel, T. Böhlke, and M. Schneider. Ecient xed point and NewtonKrylov solvers for FFT-
based homogenization of elasticity at large deformations. Computational Mechanics 54 (2014),
pp. 14971514.
[3] J. Kochmann, S. Wulngho, S. Reese, J. R. Mianroodi, and B. Svendsen. Two-scale FE-FFT
and phase-eld-based computational modeling of bulk microstructural evolution and macroscopic
material behavior. Computer Methods in Applied Mechanics and Engineering 305 (2016), pp. 89
110.
[4] J. Kochmann, S. Wulngho, L. Ehle, J. Mayer, B. Svendsen, and S. Reese. Ecient and accu-
rate two-scale FE-FFT-based prediction of the eective material behavior of elasto-viscoplastic
polycrystals. Computational Mechanics 61 (2017), pp. 751764.
[5] J. Kochmann, K. Manjunatha, C. Gierden, S. Wulngho, B. Svendsen, and S. Reese. A simple
and exible model order reduction method for FFT-based homogenization problems using a sparse
sampling technique. Computer Methods in Applied Mechanics and Engineering 347 (2018), pp. 622
638.
[6] R. A. Lebensohn, A. K. Kanjarla, and P. Eisenlohr. An elasto-viscoplastic formulation based on
fast Fourier transforms for the prediction of micromechanical elds in polycrystalline materials.
International Journal of Plasticity 32-33 (2012), pp. 5969.
[7] H. Moulinec and P. Suquet. A numerical method for computing the overall response of nonlinear
composites with complex microstructure. Computer Methods in Applied Mechanics and Engineering
157 (1998), pp. 6994.
[8] D. Peirce, R. J. Asaro, and A. Needleman. Material rate dependence and localized deformation in
crystalline solids. Acta Metallurgica 12 (1983), pp. 19511976.
[9] P. Perzyna. Thermodynamic Theory of Viscoplasticity. Advances in Applied Mechanics 11 (1971),
pp. 313354.
[10] A. Prakash and R. A. Lebensohn. Simulation of micromechanical behavior of polycrystals: nite
elements versus fast Fourier transforms. Modelling and Simulatio in Materials Science and Engi-
neering 17 (2009), p. 064010.
[11] S. Wulngho and T. Böhlke. Equivalent plastic strain gradient crystal plasticity enhanced power
law subroutine. GAMM-Mitteilungen 36 (2013), pp. 134148.
416

Large strain two-scale simulations on laptop computers

Oliver Kunc1 and Felix Fritzen∗1
1 Efficient Methods for Mechanical Analysis, Institute of Applied Mechanics (CE), University of Stuttgart
Abstract. The Concentric Interpolation method is applied to structural simulations. It is based on an

split of the input vector into amplitude and direction. The amplitude is interpolated by means of piece-
wise defined one-dimensional polynomials, while the direction is interpolated by means of spherical basis
functions. This method is well-suited for application to hyperelasticity, as the material law is only state-
dependent. Moreover, hyperelastic laws usually show distinct behavior with respect to the amplitude of the
loading and the loading “direction”, i.e. the direction of the stretch tensor in six dimensions. This very ef-
ficient evaluation of material laws renders structural two-scale simulations possible on everyday computers.
1 Introduction
Two-scale problems are known for leading to excessive computational costs if not treated efficiently.
Multiple methods for addressing this problem have been proposed in the mechanics community. Most
of them are projection based, see [3] for visco-plasticity, [2] for small strain hyperelasticity, and [1] for a
recent numerical comparison. Among these, the authors’ previous work [2] connected the projection based
reduced basis approximation of homogenized hyperelastic small strain material laws with a subsequent
kernel-based interpolation. This purely numerical second step is now extended to the finite strain context.
By focusing on provided data samples, the evaluation of hyperelastic material laws reduces to non-iterative
evaluations of elementary functions. The anticipated speed-up factors compared to the FEM are in the
order of at least 103 , while not imposing too demanding requirements on the offline phase, i.e. the amount
of data samples acting as the supporting points is manageable.
2 Problem setting
Let Ω0 be the configuration of a solid material body at rest. Under external loadings, it deforms and
translates into the region Ω, see Figure 1. The kinematic quantities X, x and u are the initial coordinates,
the deformation and the displacement, respectively. The determinant J of the gradient F = ∂x/∂X has
PSfrag repla ements Ω0

u
Ω
X x
Figure 1. Material configurations and basic kinematic quantities.
to be strictly positive in order to avoid unphysical states. Via the polar decomposition of the deformation
gradient, F = RU , the rotation tensor R and the positive definite, symmetric stretch tensor U are
obtained.
∗ Corresponding author: fritzen@mechbau.uni-stuttgart.de ORCID iD: 0000-0003-4926-0068
417
Hyperelastic material laws are based on the description of a free energy function only depending on the
deformation gradient, W = W (F ). Its first and second derivative are the first Piola-Kirchhoff stress and
the corresponding fourth-order stiffness, P = ∂W/∂F , C = ∂ 2 W/(∂F )2 , respectively.
Material objectivity implies the energy to be independent of the rotation, i.e. W (F ) = W (U ). Hence,
the stress transforms as P (RU ) = RP (U ). The components Cijkl (i, j, k, l = 1, 2, 3) of the stiffness
P3 −1
tensor transform as Cijkl (RU ) = m,n=1 δik Ulm Pmj (U ) + Rim Cmjnl (U )Rkn . Thus, it is obvious that
the material law needs to be evaluated only at the stretch tensor, and not at the deformation gradient.
This is appealing when thinking about sampling, cf. [4], as U may be represented as a six-dimensional
vector while such a representation of F requires nine dimensions.
The governing balance equation of linear momentum is DivX (P ) − b = 0, where b compounds the body
forces. Additionally, boundary conditions have to be provided. In many cases, this problem contains more
degrees of freedom than can be handled by the available computer system within the given limitations
of time. For instance, a full resolution of a fine-grained, heterogeneous, microscopic structure within a
macroscopic structural part can quickly exceed a million nodes, with three degrees of freedom each.
Asserting the separability of length scales, a two-scale framework is introduced. Therein, macroscopic
quantities are denoted by overlines, e.g. W , F , P , . . . while microscopic quantities are not highlighted,
e.g. W, F , P , . . . as before. Then, the microscopic and the macroscopic equations read
DivX (P ) = b, DivX (P ) = 0, (1)
where the micro-scale equation (1)2 is complemented by kinematic periodic fluctuation boundary condi-
tions. Among certain other kinds of admissible boundary conditions, these imply R the coupling relations
F = hF i and P = hP i, where the volume averaging operator h•i = vol(Ω0 )−1 Ω0 • dV is employed. In
the present kinematics-driven case, the macroscopic deformation gradient, F , is frequently called “the
boundary condition” of the microscopic problem. Classical two-scale schemes necessitate one solution of
the microscopic problem with the boundary condition F (X) at each macroscopic material point, X, and
for each iteration of the solution of the macroscopic problem (1)1 . After each of these solutions of (1)2 ,
the homogenized stress response is computed by means of the volume average, P = hP i.
In this work, the goal is to approximate volume averages P of solutions P to the micro-
scopic problem (1)2 very efficiently, i.e. to achieve significant speed-ups while preserving a high
level of accuracy. In other words, we seek to efficiently describe the mapping F → P . Exploiting the
transformation rule mentioned above, attention is confined to
U → P (U ). (2)
To this end, we introduce additional notation. The Hencky strain E H = log U is defined as the logarithm
of the stretch tensor. The set of Hencky strains is a vector space, i.e. arbitrary linear combinations of
Hencky strains are Hencky strains. This is in contrast to the stretch tensor, which may lose positive
definiteness under linear combinations. The stretch tensor may be recovered by means of the exponential
mapping, i.e. exp E H = U , meaning that the correspondence E H ↔ U is one-to-one.
Thus, the material law (2) may be alternatively expressed in terms of the Hencky strain. For an efficient
description, the decomposition
6
X
EH = t Ni Y (i) t = kE H k (3)
i=1
into the direction N ∈ R6 and the amplitude t is introduced. This employs the Frobenius norm k•k on
the six-dimensional Hencky space (equivalent to the space Sym(R3×3 )), spanned by a fixed orthonormal
basis Y (1) , . . . , Y (6) .
We suggest a surrogate model for the average stress P that is based on samples of Hencky
strains E H ↔ t(i) N (j) , i = 1, . . . , Namp , j = 1, . . . , Ndir , and the corresponding stresses P (exp (ti N (j) )).
If the directions are approximately uniformly distributed, then the space of Hencky strains is sampled in
a concentric manner, cf. Figure 2 and [2].
418
N (2) Y (2)
PSfrag repla ements
N (1)
(3) (1)
N Y
t(1) t(2) t(3)
N (5)
N (4)
Figure 2. Concentric sampling positions within the Hencky space, exemplary along Ndir = 5 directions and
Namp = 3 amplitudes. This visualizes the two-dimensional case.
3 Concentric Interpolation of hyperelastic material laws
Let a macroscopic stretch tensor U be given, for which the corresponding stress tensor P shall be
computed, cf. (2). We do so by separate interpolation of each component P l , l = 1, . . . , 9, of P in two
stages, based on the Concentric Interpolation scheme
P l (t, N ) ≈ w(t)> k(N ), (4)
which differs from classical kernel interpolation on hyperspheres, cf. [5], by the fact that the weight
vector, w, depends on the amplitude, t. The spherical basis functions or kernel functions
kj (N ) = exp(−γ acos2 (N · N (j) )) j = 1, . . . , Ndir (5)
employ the kernel parameter γ > 0, which needs to be determined beforehand, e.g. by a bisectional
algorithm. The univariate vector of weight functions
h i
w(t) = Pel (t, N (1) ) . . . Pel (t, N (Ndir ) ) K −1 (6)
necessitates the kernel matrix
K ij = exp(−γ acos(N (i) · N (j) )) i, j = 1, . . . , Ndir (7)
as well as radial interpolants Pel (t, N (j) ). The latter are piecewise defined polynomials that interpolate
the sampling values along each of the sampling directions N (j) . This means there is one distinct piecewise
polynomial (e.g. of degree two or three) for each sampling direction, with supporting points along its
respective sampling direction, e.g. along the radial lines in Figure 2.
Hence, the Concentric Interpolation scheme is an extension of the classical interpolation
with spherical basis functions on hyperspheres SD−1 to the surrounding Euclidean space RD .
In the present case, where one operates on the Hencky strain, the dimension is D = 6, but the scheme is
generally applicable to any D ≥ 2. In finite strain mechanics, the Concentric Interpolation method gains
significant efficiency from the fact that it is sufficient to sample U ↔ E H instead of F . This affects both
the offline phase (when the data at the supporting points is generated) and the online phase (when the
interpolation is performed).
The compatibility of the interpolant with the actual hyperelastic material law, P = ∂W /∂F , is only
guaranteed on the supporting points, e.g. on the sampling data. It is expected that the errors due to the
interpolation lie within acceptable bounds. Also, the treatment of the fourth-order stiffness, C, requires
to be investigated.
First results are promising and actual two-scale simulations of realistic complexity, e.g. with more than
105 degrees of freedom on the macroscopic scale, are anticipated in the near future.
419
Acknowledgments
This research was funded by Deutsche Forschungsgemeinschaft (DFG) within the Emmy–Noether pro-
gramm under grant DFG-FR2702/6.
References
[1] Benjamin Brands, Denis Davydov, Julia Mergheim, and Paul Steinmann. Reduced-Order Mod-
elling and Homogenisation in Magneto-Mechanics: A Numerical Comparison of Established Hyper-
Reduction Methods. Mathematical and Computational Applications 24.1 (2019). doi: 10 . 3390 /
mca24010020.
[2] Felix Fritzen and Oliver Kunc. Two-stage data-driven homogenization for nonlinear solids using a
reduced order model. European Journal of Mechanics A / Solids 69 (2018), pp. 201 –220. doi:
10.1016/j.euromechsol.2017.11.007.
[3] Felix Fritzen and Matthias Leuschner. Reduced basis hybrid computational homogenization based
on a mixed incremental formulation. Computer Methods in Applied Mechanics and Engineering 260
(2013). yig, pp. 143–154. doi: 10.1016/j.cma.2013.03.007.
[4] Oliver Kunc and Felix Fritzen. Finite Strain Homogenization Using a Reduced Basis and Efficient
Sampling. Mathematical and Computational Applications 24.2 (2019). doi: 10.3390/mca24020056.
[5] Francis J Narcowich, Xingping Sun, Joseph D Ward, and Holger Wendland. Direct and inverse
Sobolev error estimates for scattered data interpolation via spherical basis functions. Foundations of
Computational Mathematics 7.3 (2007), pp. 369–390. doi: 10.1007/s10208-005-0197-7.
420

Thermal Homogenization Using Image Data

1
Julian Lißner ∗
and Felix Fritzen1
1 University of Stuttgart, Institute of Applied Mechanics (CE)
Abstract. An image based prediction of the effective heat conductivity for highly heterogeneous microstruc-
tured materials is presented. The synthetic materials under consideration show different inclusion morphology,
orientation, volume fraction and topology. The prediction of the effective property is made exclusively based
on image data with the main emphasis being put on the 2-point spatial correlation function. This task is
implemented using both unsupervised and supervised machine learning methods. First, a snapshot proper
orthogonal decomposition (POD) is used to analyze big sets of random microstructures and, thereafter, to
compress significant characteristics of the microstructure into a low-dimensional feature vector. In order to
manage the related amount of data and computations, incremental snapshot POD methods are proposed.
In the second step, the obtained feature vector is used to predict the effective material property by using
feedforward neural networks.
1 Introduction
In material analysis and design of heterogeneous materials, multiscale modeling can be used for the
discovery of microstructured materials with tuned properties for engineering applications. However, the
discovery of materials with the desired material property, being characterized by the microstructure of
the solid, constitutes a highly challenging inverse problem. The basis for all multiscale models and
simulations is information on the microstructure and on the microscale material behavior. If at hand,
physical experiments can be replaced by – often costly – computations [4, 6] in order to determine the
material properties by virtual testing. Separation of structural and microstructural length scales can
often be assumed. This enables the use of the representative volume element (RVE) [8], characterizing
the highly heterogeneous material using a single frame (or image) of the microstructured material, on
which analytical or numerical computations can be conducted.
Supposing that similar images of the microstructured material yield similar effective material properties,
the problem is rendered suitable for machine learning (ML) models. Although ML models usually focus
on one specific task, it does not mean that the microstructure variations must be overly confined, for
instance, inclusion volume fractions ranging from 20 up to 80% are considered. The main task, thus,
persists in finding low-dimensional parameterizations of the images that capture the relevant information,
use these parameterizations to compress the image information and build a surrogate model operating
only on the reduced representation. The sampling of the data, the feature extraction and the training
of the ML algorithm constitute the offline phase in which the surrogate model is built. Typically,
the evaluation of the surrogate model can be realized almost in real-time, thereby enabling previously
infeasible applications in microstructure tailoring, interactive user interfaces and computations on mobile
devices.
Similarly to key ideas of the material knowledge system (MKS) approach [3], a reduced basis is deployed
to reduce the dimensionality of the microstructural features contained in the 2-point correlation functions.
With the sheer amount of samples required, the necessity for an incremental reduced basis computation
scheme arises. With the use of synthetic microstructure data, the costly training of the reduced basis
and of the artificial neural network [1] become feasible, enabling the creation of a surrogate model for the
image-property linkage [5].
∗ Corresponding author: lissner@mechbau.uni-stuttgart.de ORCID iD: 0000-0002-2286-5211
421
2 Microstructure classification
The underlying bi-phasic microstructure is defined by the representative volume element (RVE), which is
one periodic frame (or image) characterizing the heterogeneous material under considerations (Figure 1).
The bi phasic material consists of the matrix phase of the material and insulating inclusions. The phase
volume fraction, i.e. the volumetric part of each phase, suffices to describe the microstructure in a
homogenized setting. However, figure 2 shows, at a constant phase volume fraction, variations beyond
50% for the effective material property, the heat conductivity κ̄, are present for the underlying material.
Since additional microstructural features, e.g. geometry, are characterized by randomness, a descriptor
for the microstructural features is required. The 2-point spatial correlation function (2PCF) [7] induce a
probability distribution describing the underlying microstructure. Roughly described, each value of the
2PCF is obtained by placing a line of a certain direction uniformly randomly distributed and computing
the mean probability of finding a prescribed phase at the start and end point of the line, respectively.
Figure 1. Depicting exemplary microstructures with their respective 2-point spatial correlation functions below.
Examples of the 2PCF (Figure 1) with the corresponding RVE above show similarities between the
majorly different microstructures. Some observed characteristics of the 2PCF are:
• the maximum occurs at the corner of the domain
• preferred directions of the inclusion placement and/or orientation are captured
• the domain around the corners partially reflect the average inclusion shape
Hence, some similarities are found, particularly with respect to shape of the 2PCF at the corners and
in the center. These observations hint at the existence of a low-dimensional parameterization of relevant
microstructural features.
κ11 κ22
0.8 min/max 0.8 min/max
mean mean

m·K
m·K
0.6 0.6
W
W

κ22
κ11
0.4 0.4
0.2 0.2
0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8
12
fb [-] 12
fb [-]
Figure 2. The range of κ̄11 and κ̄22 for 15000 RVEs for the microstructures under consideration is shown. Only
1000 values are depicted.
422
3 Incremental snapshot POD of the 2PCF
The properties of the 2PCF make them an excellent choice for a snapshot proper orthogonal decomposition
(POD), constituting the unsupervised learning phase. The snapshot POD is used to construct a reduced
basis (RB) [2] that provides an optimal subspace for approximating a given snapshot matrix. The
snapshot matrix is used to compute the RB and consists of multiple individual snapshots containing the
unreduced field information, i.e. the 2PCF. The RB B is the discrete representation of a comparatively
small subspace, which minimizes the relative projection error of the given snapshots. Since the considered
microstructures allow for vastly different geometries, a humongous snapshots matrix would be required
to ensure that the RB captures all the possibilities of microstructures. Due to storage and memory
limitations this is obviously not possible and an iterative update scheme is required (Figure 3).
Pδ < ϵ
initial
basis B reduced projection
basis B error Pδ
star
ct
oje
pr
t
adjust Pδ ≥ ϵ
new snapshot
nδ = 0 compute nδ = na add snap-
update shot nδ ++
nδ < na
Figure 3. Iterative update scheme of the basis enrichment.
Assuming that a high projection error of a snapshot onto the RB implies a poor knowledge of the RB
about the microstructural features contained in the snapshot, this snapshot is a valid candidate for the
RB update. After enough candidate snapshots have been accumulated within a temporary buffer, the RB
is updated. The iterative scheme is terminated once the RB is sufficiently accurate. The resulting RB B
contains a few salient eigenmodes, containing the most reocurrent information of the given snapshots it
was computed with.
4 Data-driven Surrogate
The information contained within each RVE is given in voxel data, which consists of 400 × 400 voxels
in this study. In order to have a low dimensional descriptor of the microstructure we start with the
phase volume fraction fb of the inclusion phase, where some correlation between fb and the effective heat
conductivity can be observed (Figure 2). The RB is used to extract the feature vector of each RVE,
which is done for each snapshot si containing the 2PCF rewritten into vector format after shifting and
rescaling. The input for the regression model ξ is consequently given by

fb
ξ= ∈ R(h+1)×1 , (1)
B T si
where only the first h = 26 eigenmodes, containing the most reoccurent patterns of B, are used. This
allows a reduction of the dimension from 400 × 400 to just 27.
The training of the regression model constitutes the supervised learning phase. During the training
the machine is provided with the inputs and their desired outputs. Dense feedforward artificial neural
networks (ANN)[1] are a powerful machine learning tool to map given inputs to the desired output,
suitable for this nonlinear regression task from the feature vector (Equation (1)), to the effective heat
conductivity (Figure 2). Inspired the by human brain, dense feedforward ANNs propagate an input signal
through various neurons to obtain an output signal. In order to achieve the desired mapping, the ANN
is iteratively trained using precomputed data sets. During the training, the internal parameters of the
ANN are adjusted under the premise to minimize the prediction error.
423
5 Preliminary Results
After the costly training of the incremental snapshot POD and the ANN is concluded, accurate and fast
predictions can be made. The surrogate model has been evaluated on 2×7500 unseen test data, achieving
mean relative errors, corresponding to [5], well below 2%.
Table 1. Relative errors of the surrogate models prediction
error
κ11 κ22 κ11 κ22
measures
mean [%] 1.77 1.76 1.65 1.60
max [%] 11.7 14.1 11.6 10.5
6 Conclusions
A purely image based surrogate model to predict the effective heat conductivity has been proposed.
Microstructures with topological and morphological variations with volume fractions ranging from 20
to 80% are considered. In order to handle the amount of data required to build the surrogate, an
iterative snapshot POD method is deployed. The computed RB is then used to derive the extremely low
dimensional feature vector used by the ANN. For most of the 15000 test data, the proposed surrogate
model achieved remarkable relative errors well below 2%. A computational speedup of a factor of 40 has
been achieved compared to the intrinsically fast FANS solver [4]. The most costly step of the surrogate
model is the evaluation of the ANN input. Ongoing research indicates that further computational speedup
can be achieved by computation of the feature vector in FFT space, boosting the speedup from 40 to at
least 50. Nonetheless, the achieved accuracy is unaffected by these modifications.
Acknowledgments
Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s
Excellence Strategy – EXC-2075 – 390740016 and under grants FR2702/6, FR2702/8.
References
[1] Imad A Basheer and Maha Hajmeer. Artificial neural networks: fundamentals, computing, design,
and application. Journal of microbiological methods 43.1 (2000), pp. 3–31.
[2] R Chris Camphouse, James Myatt, Ryan Schmit, Mark Glauser, Julie Ausseur, Marlyn Andino,
and Ryan Wallace. A snapshot decomposition method for reduced order modeling and boundary
feedback control. In: 4th Flow control conference. 2008, p. 4195.
[3] Surya R Kalidindi. Computationally efficient, fully coupled multiscale modeling of materials phe-
nomena using calibrated localization linkages. ISRN Materials Science 2012 (2012), pp. 1–13.
[4] Matthias Leuschner and Felix Fritzen. Fourier-Accelerated Nodal Solvers (FANS) for homogenization
problems. Computational Mechanics 62.3 (2018), pp. 359–392.
[5] Julian Lißner and Felix Fritzen. Data-Driven Microstructure Property Relations. Mathematical and
Computational Applications 24.2 (2019), p. 57.
[6] Christian Miehe. Strain-driven homogenization of inelastic microstructures and composites based on
an incremental variational formulation. International Journal for Numerical Methods in Engineering
55 (2002), pp. 1285–1322.
[7] S Torquato and G Stell. Microstructure of two-phase random media. I. The n-point probability
functions. The Journal of Chemical Physics 77.4 (1982), pp. 2071–2077.
[8] Salvatore Torquato. Random heterogeneous materials: microstructure and macroscopic properties.
Vol. 16. Springer Science & Business Media, 2013.
424

Numerical and experimental characterization of addi-

tively manufactured mild steel under tension
1
Jorge Lizarazu ∗
, Luise Göbel1 , Stefan Linne1 , Susanne Kleemann1 and Tom Lahmer1,2
1 Materialforschungs- und -Prüfanstalt an der Bauhaus-Universität Weimar, 99423 Weimar, Germany.
2 Institute of Structural Mechanics, Bauhaus-University Weimar, 99423 Weimar, Germany.
Abstract. For the case of metals, Additive Manufacturing (AM), also known as 3D Printing, is a technology
based on a layer-by-layer welding procedure able to create 3D products.
In the present study, the tensile behavior of additively manufactured mild steel is studied using both exper-
imental and numerical approaches. Due to several heating and cooling cycles in the welding procedure, the
microstructure of the metallic product is strongly modified and therefore, it is not as homogeneous as the
original material. As a result, the product is first analyzed with optical and scanning electron microscopes,
in order to identify the phases and extract the grain properties. Based on this information, the microstruc-
ture of the representative volume element (RVE) is generated by means of two approaches: The first one
consists of using a Voronoi Tessellation and the second one comprises the digitization of the image obtained
from the scanning electron microscope (SEM). After the definition and assignment of the constitutive laws
to each phase, a subsequent numerical tensile test is carried out. The result of this simulation is compared
with experimental tensile tests of specimens, which were produced using the wire arc additive manufacturing
technique.
The experiments indicate that the stiffness of the additively manufactured products is similar in both parallel
and perpendicular welding directions. On the contrary, the ductility exhibits more visible differences in these
two directions. Besides, the material, shows a softening behavior in one direction.
Introduction.
The current applications of additively manufactured metals require a proper knowledge of the material;
so that, their response under several type of loading can be estimated accurately by using numerical
techniques. In pursuit of this, a study regarding uniaxial tensile behavior of 3D printed products is
necessary, in order to set up the basis for the simulation under more complex loading conditions.
In this work, the procedure already proposed by [3] is applied to 3D printed mild steel products, which
were created using the wire arc additive manufacturing (WAAM) technique.
WAAM is used to create middle to large sized products and has the following main advantages [4, 7]:
Relative high deposition rates and a higher usage efficiency of the material. However, the technique is only
able to create products of middle level of structural topology complexity with a high surface roughness;
being necessary to post-process the products in a specific way [1]. Besides, since WAAM is very sensitive
to the process parameters, building sequences and welding disturbances [4], repeatability of the products
can be an issue; therefore, an additional investment for sensing and controlling might be necessary. The
process parameters are: Wire feed speed, welding speed, layer thickness, arc voltage, arc current, wire
diameter, wire type, shielding gas type and gas flow. Table 1 summarizes the typical features of WAAM.
Table 1. Typical features of WAAM

Layer thickness Deposition rate Dimensional accuracy Surface roughness
(mm) (kg/h) (mm) (µm)
1-2 Until 6 ± 0.2 500
∗ Corresponding author: jorge.lizarazu@mfpa.de
425
1 Study of the microstructure.
In order to analize the microstructure under the light microscope (LM) and SEM, some square samples
of side 1.5 cm were cut from a plate, whose two faces were previously polished; later, the samples were
etched.
1.1 Investigation under the light microscope.
The purpose of the investigation under the LM is to identify the number and boundaries of heat affected
regions. Figure 1 shows that only one of such regions exists, being even impossible to identify the
boundaries of each welding layer.
Figure 1. Images of the material under the light microscope.
1.2 Investigation under the scanning electron microscope.
The purpose of this analysis was to identify the phases of the material and extract grain parameters.
Two phases were identified in Figure 2(a): Ferrite and Cementite in darker and lighter color respectively.
Figure 2(b) shows the existence of several pores (in red circles) and inclusions (in red squares).
(a) Identification of phases (b) Identification of pores and inclusions
Figure 2. SEM images of the microstructure.
2 Experimental Investigation.
2.1 Microtensile test.
The purpose of this test is to extract qualitative information regarding the crack formation and propa-
gation. All this process is depicted and self explained in Figure 3.
426
(a) First microstructural changes (b) Crack formation
(c) Crack opening (d) Break down
Figure 3. Formation and propagation of cracks in the material.
2.2 Macrotensile test.
Several specimens of rectangular shape (1 cm x 10 cm) and bottle neck shape (inscribed in a rectangle of
dimensions 1.5 cm x 12 cm) were tested in order to build the stress-strain relations, observe the mechanical
behavior and extract mechanical properties. The specimens were extracted from the outer boundaries of
two plates, as well as from the inside part of them; additionally, the specimens were classified as parallel
and perpendicular (to the welding direction).
In principle, no differences were noticed between the mechanical behavior of the specimens located at
the outer boundary and the inner part of the plates. It was observed that the stiffness of the specimens,
parallel and perpendicular, are similar; however, the ductility of the parallel specimens (Figure 4(d)) is
twice as big as the ductility of the perpendicular specimens (Figure 4(b)). Besides, it was noticed that
after converting the engineering measurements of stress and strain into the corresponding true values,
the stress-strain curve exhibits a softening behaviour.
3 Numerical simulation.
3.1 Polycristaline microstructure and generation of the finite element mesh.
Two approaches were considered for the generation of the microstructure. The first one based on the
Voronoi Tesselation and the second one based on the digitization of the SEM image.
3.1.1 Synthetic generation of the microstructure.
The first approach consisted on using the computer program Neper [5] to generate the microstructure. It
is based on the grain parameters already extracted from the SEM image (Figure 2(a)) and the Voronoi
Tesselation. The synthetically generated microstructure and the corresponding finite element mesh are
shown in Figure 5.
427
(a) Eng. values of perpendicular specimen (b) True values of perpendicular specimen
(c) Eng. values of parallel specimen (d) True values of parallel specimen
Figure 4. Formation and propagation of cracks in the material.
(a) Voronoi Tesselation (b) Finite element mesh
Figure 5. Synthetic microstructure and its discretization into finite elements.
3.1.2 Image scanning based generation of the microstructure.
By using the computer program OOF2, the SEM image is automatically digitized and discretized, see
Figure 6.
3.2 Constitutive relations.
After generating the microstructure and meshing each phase of the microstructure separately, the con-
stitutive behavior to each phase is specified as follows:
3.2.1 Ferrite.
For this phase, a linear elastic behaviour with parameters E = 200 GPa and µ = 0.3 is defined. Addi-
tionally, a plastic behaviour following Eq.(1) is considered, whose parameters are thoroughly explained
428

A Homogenization Method with Minimal RVE Bound-

ary Conditions
Maximilian Müller
∗ 1 and F. Gruttmann
1
1 TU Darmstadt, Fachgebiet Festkörpermechanik, Franziska-Braun-Straÿe 7, 64287 Darmstadt
Abstract. In this contribution, a homogenisation method based on the Irving Kirkwood theory is intro-
duced. From basic assumptions for the relations between micro and macro scale mass density and linear
momentum, homogenisation formulas for the stress tensor and body force vector are derived. This ap-
proach ensures compliance between the microscopic and macroscopic model without the need to full the
Hill condition a priori by means of a specic set of boundary conditions. Therefore a much wider vari-
ety of boundary conditions can be used in the representative volume element (RVE). As a rst example,
minimal boundary conditions are used. These only prevent rigid body motions. Loading of the RVE is
realised through a side condition, which relates the strains of the two scales and is introduced into the
weak form of equilibrium via Lagrange multipliers. Computations for three dimensional beam like struc-
tures serve as a rst test. A two-scale homogenisation scheme, that uses beam elements on macro level
and solid brick elements on micro level is implemented. The implementation is veried by obtaining the
homogenised stiness matrix for a homogeneous beam. Finally, structures with inhomogeneities in length
direction are studied and obtained results are compared to benchmark solutions from a full-scale model.
Introduction
The existing and in FE2 widely used approach of fulfulling the Hill-Mandel condition for homogenisation
a priori by using certain sets of boundary conditions for the representative volume element (RVE), limits
possible boundary conditions very strictly and often produces undesired boundary and other eects. The
desire to get rid of those restrictions led to the development of a more general homogenisation method
based on the Irving-Kirkwood theory [2, 3, 4]. For this contribution we apply the method to beam like
structures undergoing small strains and use of beam elements on macro level. Within the RVE there will
be no other boundary conditions than the ones necessary to restrict rigid body movements. The loading
of the RVE will be realised with an additional constraint. For numerical examples, we study a beam with
a square cross section and periodic holes perpendicular to the length direction.
1.1 Homogenization Theory
In order to describe a body inhomogeneities, which are much smaller than its macroscopic measure-
ments, the heterogenous body will be described in two scales. We will assume homogeneous properties
on a macroscopic scale and take account of the inhomogenieties on a microscopic scale. The homogenous
properties on the macroscopic level are then derived from the microscopic level by means of homogeniza-
tion.
In detail, there shall be given a body B inhabiting the volume R with boundary ∂R. A point in B shall
T
be described with coordinates y = [y1 , y2 , y3 ] dening the macro scale. For every point y there is an
assigned surrounding volume P m of the micro scale, described with coordinates x = [x1 , x2 , x3 ] . T
For all subdivisions of R and all P m the mass and linear momentum balance laws hold:
ρ̇α + ρα ∇y · (v α ) = 0, (1)
α α α α
α
ρ v̇ = ∇x · σ + ρ b . (2)
∗ Corresponding author: mmueller@mechanik.tu-darmstadt.de
429
Here, α is either m for microscopic or M for macroscopic, indicating the scale in which the quantity is
dened. Further, ∇x · (·) denotes the divergence of (·) with respect to x.
These balance laws for the two scales are not independent as they describe the same body. So we postulate
the following homogenization laws for mass density und linear momentum in accordance with [2]:
Z
ρM (y, t) = ρm (x, t)g(y, x) dV m , (3)
R
Z
ρM (y, t)v M (y, t) = ρm (x, t)v m (x, t)g(y, x) dV m . (4)
R
In (3) and (4) g(y, x) is a weighting function, which needs to fulll a few basic properties in order to not
compromise the theory. For further information on those properties and their implications see [2]. There,
it is also shown, that the postulated homogenization laws are consistent with the balance laws (3) and
(4) and lead to additional homogenization laws for the body force vector and stress tensor. For the quasi
static case discussed here, they are:
Z Z
ρM bM := ρm b m g d V m , (5) σ M := σ m g dV m . (6)
R R
For this contribution we will set the weighting function g(y, x) = V1m , where V m is the volume of the
RVE. Also, as stated before, deformations will be limited to small strains.
1.2 Variational Formulation
With the homogenization laws for the stress tensor (5) and body force vector (6) the transition from
micro to macro level is complete. We now need a transition law from macro to micro. Looking at the
homogenization laws, it appears naturally to use the same homogenizational connection for the micro
and macro strains:
Z Z Z
M m m 1 m m
ε = ε g dV = m ε dV ⇔ (εM − εm ) dV m = 0. (7)
R V R R
With given macro strains εM , equation (7) will be implemented as an additional constraint into the weak
form of the micro problem via Lagrange multipliers. For each of the n components of the strain vector,
we will need one Lagrange multiplier, which will be arranged in a vector λ̄ = [λ̄1 , · · · , λ̄n ].
M
Variation with respect to displacements u, Lagrange multipliers λ̄ and macro strains ε leads with
vanishing body forces to the following weak form of the problem:
Z Z Z
T T
δεmT σ m dV m + δ λ̄ (εM − εm ) dV m + λ̄ (δεM − δεm ) dV m = 0. (8)
R R R
The variation with respect to εM is not necessary but leads to a simpler implementation of the homoge-
nization scheme, that can use existing code.
1.3 Application to Beams
We start with linear Timoshenko beam kinematics for the displacements ū = [ūx , ūy , ūz ]T of an arbitrary
point with coordinates (x, y, z) where the coordinate x is aligned to the beam axis:
ūx (x, y, z) = ux (x) + βy (x)z − βz (x)y (9)
ūy (x, y, z) = uy (x) − βx (x)z (10)
ūz (x, y, z) = uz (x) + βx (x)y (11)
Here, ux , uy , uz denote the displacements of the beam axis and βx , βy , βz are rotations along their corre-
sponding axis.
The linearized Green strains εv = εM are then given by
   0 
εx ux + βy0 · z − βz0 · y
 εy   0 
   
 εz   0 

εv =   
= 0 . (12)
2ε  u − β − β 0
· z 
 xy   y z x 
2εxz   uz + βy + βx · y 
0 0
2εyz v 0
430
They can be associated to the given macroscopic beam strains εb :

        
εx u0x εx 1 0 0 0 z −y εx
γxy  u0y − βz   εy  0 0 0 0 0 0  γxy 
   0      
 γxz  uz + βy   εz  0 0 0 0 0 0   
 
εb :=   :=    ⇒ 
εv =   = P b εb =    γxz  . (13)
0   0 −z 0 
 
 ϑx   βx0  2εxy   1 0 0  ϑx 
 κy   βy  2εxz  0 0 1 y 0 0   κy 
κz b βz0 2εyz v 0 0 0 0 0 0 κz b
In equation (8) the Lagrange multipliers can be indentied as a stress-like quantity. While on macroscopic
level they need to represent the stress resultants σ b = [N, Qy , Qz , Mx , My , Mz ]T , on microscopic level
they need to represent the cauchy stresses σ v = [σx , σy , σz , τxy , τxz , τyz ]T . They can be transformed via
the following equation:
    
λσx 1 0 0 0 z −y λ̄N
 λσy  0 0 0 0 0 0   
     λ̄Qy 
 λ σz  0 0 0 0 0 0   λ̄Qz 
λ=  
λτxy  = P λ λ̄ = 0 f1 (y)
  (14)
   0 −z 0 0  
 λ̄Mx 

λτxz  0 0 f2 (z) y 0 0  λ̄My 
λτyz 0 0 0 0 0 0 λ̄Mz
While the Lagrange multipliers λ̄, which correspond to the macroscopic stress resultants of the beam,
are constants as they are constant for each Gauss point of the RVE, the altered multipliers λ (and
correspondingly σ v ) should not be constant over the cross section.
As in basic beam theory, the normal stress σx is linear in z and y with the gradient being dependent from
the bending moments. The shear stresses τxy and τxz however, are assumed constant in basic beam theory
which leads to a too sti reaction, especially in cases like sandwich beams with a very high dierence in
stiness between core and covering layers. In order to circumvent this problem we assume much more
realistic quadratic shear stresses. For this contribution we use the well known formula for square cross
sections and set

3 y2 3 z2
f1 (y) = 1−4 2 and f2 (z) = 1−4 2 . (15)
2A b 2A h
Wherein b and h are the width and height of the cross section and A = bh is the area of the cross section.
Inserting equations (13) and (14) into the weak form (8) gives:
Z Z Z
δεmT σ m dV m + δλT P Tλ (P b εM
b − εm
) dV m
+ λT P T M m
λ (P b δεb − δε ) dV
m
= 0. (16)
R R R
1.4 Finite Element Formulation
For the macroscopic level, standard 3D beam elements with six degrees of freedom (three translations,
three rotations) are used. On microscopic level we use 3D brick elements. A Linearization of (16) and a
standard iso-parametric FE-formulation leads to:
L[. . . ] = δV mT (K∆V m + A∆λ + F ) + δλT (AT ∆V m + T ∆εM ) + δεMT (T T ∆λ) = 0. (17)
In which m = BV m , with V m being the displacement vector, was used and
Z Z Z
K := B T DB dV m , D=
∂σ
, A := − B T P λ dV m , T := T
P λ P b dV m .
R ∂ε R R
We also redened εM := εM b for better readability.

Next, we follow the standard homogenization scheme presented in [1], which can be used here in a very
similiar, straight forward way.
431
2 Numerical Examples
2.1 Verication on a Homogeneous, Linear Elastic Cantilever
First, we want to verify the model on a homogeneous beam with a square cross section and linear elastic
material. We use a beam with length of 16cm and height and width of 1cm. With our improvements to
the shear stresses, the entries of the homogenized stiness matrix converge towards the expected values
D M = diag{EA, κy GA, κz GA, GIp , EIy , EIz }. With E being the Youngs modulus, G the shear modulus,
A the area of the cross section, Iy , Iz and Ip = Iy + Iz the area moments of inertia and κy = κz = 56 the
shear correction factors.
2.2 Cantilever with Plasticity and Holes
We now investigate a beam with the same dimensions as before, but with periodic holes in y -direction
with a diameter of 0.25cm. We consider a linear elastic, plastic material with linear isotropic hardening
(E = 100000 cmkN , ν = 0, y = 600 kN , ξ = 10000 kN ). The cantilever is loaded at the free end with a
2 0 cm2 h cm2
single force F in z-direction.
For a reference solution we use a full scale model with linear 3D solid elements, which consists of 16 holes
over a beam length of 16cm. For our multiscale solution, we use ten beam elements on macro level and
a cubic RVE with side length of 1cm and one hole, g. 1. As a result, we compare the applied force F
over displacement uz of the free end for both models. We then see a perfect match for the linear elastic
part and a good agreement for the plastic part of the structural behaviour (g. 2).
12,5
10
force F [kN]
7,5
2,5 reference
multi scale
0
0 0,5 1 1,5 2 2,5 3
displacement u z [cm]
Figure 1. RVE for a beam with holes Figure 2. Load-displacement diagram of the cantilever
References
[1] F. Gruttmann and W. Wagner. A coupled two-scale shell model with applications to layered struc-
tures. International Journal for Numerical Methods in Engineering 94 (2013), p. 12331254.
[2] K. K. Mandapu, A. Sengupta, and P. Papadopoulos. A homogenization method for thermomechanical
continua using extensive physical quantities. Proceedings of the Royal Society 468 (2012), p. 1696
1715.
[3] B. S. Mercer, K. K. Mandapu, and P. Papadopoulos. Novel Formulations of Microscopic Boundary-
Value Problems in Continuous Multiscale Finite Element Methods. Computer Methods in Applied
Mechanics and Engineering 286 (2015), p. 268292.
[4] M. Müller and F. Gruttmann. A homogenisation method based on the Irving-Kirkwood-Theory.
Proceedings of the 7th GACM Colloquium on Computational Mechanics (2017). url: http://dx.
doi.org/10.18419/opus-9334.
432

Cauchy-Born rule based continuum simulation of the

mechanical properties of single-layered graphene sheets
Qige Zheng∗ 1 and Jens Wackerfuÿ1

1 Institute of Structural Analysis, University of Kassel, Kassel, Germany
Abstract. Besides experiments, numerical simulations are also necessary for studying the mechanical properties
of graphene sheets. In this paper, the Cauchy-Born rule is used in the formulation of the continuum mechanical
model. According to the original Cauchy-Born rule, the positions of the atoms within the crystal lattice follow
the overall strain of this crystalline solid in the case of small deformation. However, it is restricted to three-
dimensional crystalline materials or the in-plane deformations of a two-dimensional material (like graphene).
The exponential Cauchy-Born rule or the higher order Cauchy-Born rule, which are modied forms of the
original and applicable to shell like structures, can be used to describe the out-of-plane deformations of a
single-layered graphene.
Introduction
Due to the exceptional features of graphene sheets, many research projects are focused on them since they
were discovered in 2004 [10]. Besides experiments, also numerical simulations are necessary for studying
their mechanical properties. The models which are used in the simulations can be classied into two
categories: molecular mechanical models and continuum mechanical models.
Eects on the atomic scale can be described explicitly with molecular mechanical models. However,
this type of models is restricted to simulate structures of small sizes. For determination of the mechan-
ical behaviors of technical relevant large graphene nanostructures at the macroscopic level, it can be
calculated eciently with continuum mechanical models. Some classical formulations of these models are
able to characterize the small strain response of graphene sheets through elastic moduli such as Young's
modulus, Poisson's ratio [6, 8]. The continuum mechanical models can also be derived from interatomic
potentials describing the energy and the interactions within an atomistic system [1, 5]. In this way, the
underlying atomistic structure is considered in the constitutive relation which is a priori valid for the
scenarios of large deformations.
In this paper, a 4-node quadrilateral element with a Cauchy-Born rule based constitutive relation is
formulated for studying the mechanical properties of a single-layered graphene sheet.
1 Continuum model
In this section, a continuum model is formulated based on the quasicontinuum (QC) method [12] for
solving the static problems of graphene with absolute-zero temperature (0 K). In context of the QC
method, the internal energy E of the whole atomistic system is equal to the strain energy which is only
related to the interatomic potentials U in the situation of 0 K. And the strain energy density W is
expressed as
nc
1 X
W = Ui , (1)
Ac i=1
∗ Corresponding author: zheng@uni-kassel.de
433
where Ui and nc denote the interatomic potential on atom i and the number of atoms in the se-
lected domain (so-called representative cell) with radius Rc in the atomistic structure (see Fig. 1 (c)).
Ac = nc · Aatom is the area of the representative
√
cell. Aatom is the area of the representative cell with only
one atom and given as Aatom = 3 4 3 rb 2 with rb denoting here the equilibrium bond length of undeformed
graphene (see Fig. 1 (b)). In context of graphene, Ui can be the Brenner potential [3], the DREIDING
potential [9] or the modied Morse potential [2].
Therefore, the internal energy E of the whole atomistic system and the total energy π are formulated as
Z n
X
E= x, F ) dA0 ,
W (x π=E− f̄f i · u i . (2)
Ω0 i=1
where f̄f i and u i are the external load applied to the atom i and the displacement of this atom. Ω0
and dA0 denote the domain occupied by graphene and surface element in initial conguration. n is the
number of atoms in the whole system. Following the original Cauchy-Born rule [4], each hexagon cell of
graphene is deformed according to the local continuum deformation gradient (see Fig. 2) in the situation
of in-plane deformation. This relationship is expressed as
x
∂x
r = F ·R with F = , (3)
X
∂X
where R is a lattice vector in the undeformed conguration and r is the corresponding lattice vector in
the deformed conguration. F represents the local deformation gradient. x and X denote the positions
of the atoms in undeformed and deformed conguration.
The equilibrium deformation is identied through the energy minimization process of π , as

Xn
∂E
δπ = ( − f¯i ) · δu
ui = 0 . (4)
ui
∂u
i=1
According to the QC method, nr atoms are selected as representative atoms from the atomistic struc-
ture. The displacements and current positions of other atoms are approximated by functions of the
displacements and positions of the representative atoms. Consequently, the Eq. (4) can be rewritten as
Xnr
∂E
δπ h = ( − f¯i ) · δu
ui = 0 , (5)
ui
∂u
i=1
where π h is the approximate total energy. i denotes here the indexes of representative atoms. The Eq. (5)
is equivalent to a closed system of 3 · nr equations with 3 · nr variables. u i can be calculated through
solving this system of equations with Dirichlet boundary conditions.
η
y
x
Figure 1. Schematic diagram of: (a) A 4-node quadrilateral plate/shell element (red) with underlying atomistic
structure, (b) representative cell corresponding to an atom, (c) representative cell with radius Rc .
With regard to the EulerLagrange equation [7], the denition of rst Piola-Kirchho stress P and
Eq. (2), ∂E
u can be transformed as following
∂u
Z Z
∂E ∂E ∂E ∂W
= = · ∇0 = · ∇0 dA0 = P · ∇0 dA0 . (6)
u
∂u x
∂x F
∂F F
Ω0 ∂F Ω0
434
∇0 denotes the gradient operator in undeformed conguration. The Eq. (5) is consequently rewritten as
nr Z
X
( P i · ∇0 dA0 − f¯i ) · δu
ui = 0 , (7)
i=1 Ω0
which represents the equilibrium equations for the graphene in the situation of in-plane deformation based
on QC method here.
r F R
Figure 2. Schematic illustration of the original Cauchy-Born rule
Graphene is a crystalline lm with the thickness of one atom. In order to consider also out-of-plane
deformations of graphene, the curvature eects have to be taken into account. Therefore, a higher order
Cauchy-Born rule
1
r = F · R + G : R ⊗ R with G = ∇0F (8)
2
is applied [11]. The tensor G is the second-order deformation gradient. With Eq. (8) the corresponding
equilibrium equations are given by
nr Z
X ∂W
( P i · ∇0 + Q i : ∇0 ∇0 dA0 − f¯i ) · δu
ui = 0 with Q i := . (9)
Ω0 Gi
∂G
i=1
2 Finite element (FE) formulation
In the FE approximation, a 4-node nite element is formulated (see Fig. 1 (a)). The mapping from the
parameter space into the physical space and the approximation of displacement are written in the usual
way, namely
X4 4
X
xh = xI , u h =
NI (ξ, η)x uI ,
NI (ξ, η)u (10)
I=1 I=1
where I = 1, · · · , 4 are the nodes in an element and u I = [uI , vI , wI ]T are the nodal DOFs. The
undeformed conguration Ω0 is partitioned into nite elements Ωe0 with e = 1, · · · , ne , where ne is the
number of elements. Consequently, the total internal energy in Eq. (2) can be rewritten with regard to
Gaussian quadrature rule as
ne Z
X nq
ne X
X
E= x) dA0 ≈
W (xe
ωqe W e (x
x(ξq , ηq )) , (11)
e=1 Ωe0 e=1 q=1
where ωqe and (ξq , ηq ) are the quadrature weights and points for element e.
The system of equilibrium equations (Eq. (7) or (9)) is solved by applying the Newton-Raphson method.
For the in-plane deformation, the global residual vector R I and stiness matrix K IJ follow as
X n Z
∂π h e
RI = − f¯I = P e · ∇0 NI dA0 − f¯I , (12)

uI
∂u Ωe0
e=1
X n Z
∂ 2 πh e
∂2W e
K IJ = = C e : (∇0 NI ⊗ ∇0 NJ ) dA0 with C e = , (13)
uI ∂u
∂u uJ Ωe0 F e ∂F
∂F Fe
e=1
435
where I and J are node indexes. For the out-of-plane deformation, the corresponding global residual
vector and stiness matrix are
Xne Z
∂π h ¯
RI = −fI = P e · ∇0 NI + Q e : ∇0 (∇0 NI ) dA0 − f¯I , (14)
uI
∂u e
e=1 Ω0
Xne Z
∂ 2 πh
K IJ = = C e1 : (∇0 NI ⊗ ∇0 NJ ) + C e2 :̇(∇0 NI ⊗ ∇0 (∇0 NJ ))
∂uuI ∂uuJ Ωe
e=1 0
+ C e3 :̇(∇0 (∇0 NI ) ⊗ ∇0 NJ ) + ∇0 (∇0 NI ) : C e4 : ∇0 (∇0 NJ ) dA0 (15)

2 e 2 e 2 e 2 e
∂ W ∂ W ∂ W ∂ W
with C e1 = , C e2 = , C e3 = , C e4 = .
F e ∂F
∂F Fe F e ∂G
∂F Ge Ge ∂F
∂G Fe Ge ∂G
∂G Ge
Conclusions
In this paper, a 4-node quadrilateral element with a Cauchy-Born rule based constitutive relation is
presented and applied in context of the QC method for studying the mechanical properties of graphene.
Compared with full atomistic simulation, this model has much less DOFs. Its constitutive relation is
derived from interatomic potentials describing the energy and the interactions within an atomistic system.
Therefore, it is more suitable for the scenarios of large deformations than other models based on linear
elastic hypothesis. But this model is found on the precondition that the underlying atomistic structure
is subjected to homogeneous deformations, i.e., non-homogeneous deformations can not be modelled.
So this model should be combined with full atomistic simulation for determination of the mechanical
behaviors of technical relevant large graphene nanostructures.
References
[1] M. Arroyo and T. Belytschko. Finite crystal elasticity of carbon nanotubes based on the exponential
Cauchy-Born rule. Phys. Rev. B 69 (2004), p. 115415.
[2] T. Belytschko, S. P. Xiao, G. C. Schatz, and R. S. Ruo. Atomistic simulations of nanotube fracture.
Phys. Rev. B 65 (23 2002), p. 235430.
[3] Donald W. Brenner. Empirical potential for hydrocarbons for use in simulating the chemical vapor
deposition of diamond lms. Phys. Rev. B 42 (15 1990), pp. 94589471.
[4] J.L. Ericksen. On the CauchyBorn Rule. Mathematics and Mechanics of Solids 13.3-4 (2008),
pp. 199220.
[5] C. Findeisen and J. Wackerfuÿ. A general approximation of the exponential CauchyBorn hypothe-
sis to model arbitrarily shaped shell-like nanostructures within continuum mechanics. International
Journal for Numerical Methods in Engineering 105.10 (2016), pp. 747780.
[6] S. Govindjee and J. L. Sackman. On the use of continuum mechanics to estimate the properties of
nanotubes. Solid State Communications 110.4 (1999), pp. 227 230.
[7] H. Hardy. Euler-Lagrange Elasticity: Dierential Equations for Elasticity without Stress or Strain.
Journal of Applied Mathematics and Physics 1 (2013), pp. 2630.
[8] Jian Ping Lu. Elastic Properties of Carbon Nanotubes and Nanoropes. Phys. Rev. Lett. 79 (1997),
pp. 12971300.
[9] Stephen L. Mayo, Barry D. Olafson, and William A. Goddard. DREIDING: a generic force eld
for molecular simulations. The Journal of Physical Chemistry 94.26 (1990), pp. 88978909.
[10] K. S. Novoselov, A. K. Geim, S. V. Morozov, D. Jiang, Y. Zhang, S. V. Dubonos, I. V. Grigorieva,
and A. A. Firsov. Electric Field Eect in Atomically Thin Carbon Films. Science 306.5696 (2004),
pp. 666669.
[11] R. Sunyk and P. Steinmann. On higher gradients in continuum-atomistic modelling. International
40.24 (2003), pp. 6877 6896. issn: 0020-7683.
Journal of Solids and Structures
[12] E. B. Tadmor, M. Ortiz, and R. Phillips. Quasicontinuum analysis of defects in solids. Philosophical
Magazine A 73.6 (1996), pp. 15291563.
436

Shear Deformations in a Multiscale Shell Formulation

Jan Zoller ∗ 1 and Friedrich Gruttmann1
1 Technische Universität Darmstadt, Fachgebiet Festkörpermechanik, Franziska-Braun-Straÿe 7, 64287 Darmstadt
Abstract. In this work a two-scale shell model is introduced based on the algorithm in [1], which carries out the
concept of the FE2 -method. The representative volume element (RVE) is discretised with solid elements. The
application of global strains to the RVE via displacement boundary conditions leads to an incorrect, and with
increasing in plane size of the RVE even vanishing, shear stiness. The reason is that the RVE reacts to shear
deformations not only with shear stresses but also with bending stresses. Therefore, the theory in [2] for beams
is transferred to the shell problem. In order to eliminate that resulting fallacious bending moment a constraint
on the normal stress distribution is implemented. The use of periodic boundary conditions allows warping of the
cross section. Thus boundary eects can be reduced or even avoided in the RVE. However, an introduced shear
deformation together with permissible periodic displacements leads to a rigid body movement and, therefore,
no shear stresses and stiness are obtained. Hence, it is necessary to introduce a further constraint. To
consider inhomogeneities of the material, the constraint is established with respect to the shape of the normal
stress over the cross section. Both constraints are incorporated using the method of Lagrange multipliers.
Introduction
Shell elements using Reissner-Mindlin kinematics are commonly used for examining large shell structures,
because they describe the kinematic behavior quite simply, and thus with few degrees of freedom. How-
ever, this assumption is not sucient to adequately describe the deformation over the thickness direction
for complex material composites.
By using the multiscale method, it is still possible to use these simple shell elements on the classical global
scale. A detailed behavior of an arbitrary material distribution and structure is taken into account on an
additional local scale by investigating a unit cell with a representative material composition, a so-called
representative volume element (RVE). Within this RVE local eects can also be considered. This enables
the calculation of large and complex structures. Further, by discretizing the unit cell with solid elements,
it is possible to use existing algorithms for solid elements for the calculation in which the advantages of
shell elements should be utilized. The two scales are coupled by deforming the RVE according to the
shell strain from the associated global Gauss point p and from this, the shell stress resultants σ and
linearized stress resultants D, with the shear stiness record Dγ , are calculated.
T  ε 
= ε11 ε22 ε12 κ11 κ22 κ12 γ13 γ23 (1) D Dεκ 0
T D = Dκε Dκ 0  (3)
n n
σ = 11 22 12 n m 11 m 22 m 12 q 13 q23 (2) 0 0 Dγ
1 Homogenization scheme
An algorithm for a two-scale shell formulation can already be found in [1], which is the basis for the
presented model. The local boundary value problem has to be subdivided into constrained (vb ) and
unconstrained (va ) degrees of freedom
N
X
1 1 δva ga kaa kab ∆va
δVpT (K∆V + G)p = · + =0. (4)
ARV E ARV E δvb e gb e kba kbb e
∆vb e
e=1
∗ Corresponding author: zoller@mechanik.tu-darmstadt.de
437
For displacement boundary conditions (DBC) the applied displacements ūI of a constrained node I are
calculated by means of the linear map AI from the global strains
ūI = AI , δūI = AI δ , ∆ūI = AI ∆ . (5)
These equations are also applied for periodic boundary conditions (PBC), which here are newly introduced
for a two-scale shell formulation. This then leads to the vector vb including all degrees of freedom which
are involved in the PBC. In preparation for a later constraint, the shear distortion is brought into action
by applying a deformation in thickness direction for PBC. This requires the consideration of the 3rd row
of AI and thus AI diers for DBC and PBC
DBC: PBC:
 1 1 
" 1 1 # x 0 2y xz 0 2 yz 0 0
x 0 2y xz 0 2 yz z 0  
AI = 1 1 , AI =  0 y 1
2x 0 yz 1
2 xz 0 0 . (6)
0 y 2x 0 yz 2 xz 0 z
0 0 0 − 12 x2 − 21 y 2 − 21 xy z z
For the boundary value problem, all lines from AI belonging to the constrained degrees of freedom vb are
combined in the matrix Ae . The Matrix ae is an incidence matrix. By making use of a static condensation
process, the global shell stress resultants σ and the global linearized stress resultants D are obtained from
the local boundary value problem
1 1
σ= Gb − LT K−1 Ga , D= M − LT K−1 L with (7)
ARV E ARV E
N
[ N
X N
X N
X N
X
K= aTe kaa ae , L= aTe kab Ae , M= ATe kbb Ae , Ga = aTe ga , Gb = ATe gb . (8)
e=1 e=1 e=1 e=1 e=1
2 Eliminating the fallacious bending moments
The procedure in [1] leads to correct membrane

and bending stinesses. However, the shear sti- 1.2
[−]
ness decreases with increasing RVE in-plane sizes, 1

as can be seen in Fig. 1 (DBC) for the investigation DBC
/D11,Ref.
0.8
of a homogeneous RVE. This eect corresponds to DBC+ C1
0.6
PBC
γ
the behavior described in [3] for a two-scale beam 0.4

formulation. The decrease of the shear stiness PBC+C1+C2
D11,i
0.2
results from the fact that the RVE reacts to the
γ
0
initialized shear deformation with both shear and
012 4 8 16 32
bending stresses, thus a wave-like deformation g-
ure is established, as illustrated in Fig. 2. This RVE length to height [−]
can be explained by a simplied beam model, for
which the two in-plane dimensions are examined Figure 1. Relative shear stiness versus in-plane size of
separately, α ∈ {x, y}, Fig. 3. a homogeneous RVE.
To prevent this, the fallacious linear moment needs to be eliminated. For this purpose, the bending
stresses which conform to the linear moment are set to zero. The shape functions for the Lagrange
multipliers λαi and µαi correspond to the moment curve in Fig. 3. The constraint has to be implemented
considering all normal stresses σi with i ∈ {x, y, z}, otherwise, because of the constitutive matrix, the
stresses are shifted to another direction and the wave-like deformation remains
Z
σi · (zλαi + µαi ) dV = 0 with i ∈ {x, y, z}, α ∈ {x, y}. (9)
VRV E
The inuence of boundary eects disappears with increasing RVE dimensions, Fig. 1 (DBC+C1). This
constraint is also required for the use of PBC. The simplied beam element and thus also the shape
functions have to be adapted, Fig. 3. As a result of these constraints, there are 12 additional degrees of
freedom for the RVE.
438
DBC: PBC:
γ
L L/2
z γL
y x L/2
γ
+
z
y uz 0
x
−
Figure 3. Simplied beam models and corresponding shear
Figure 2. Wave-like deformation gure for shear deformations, shear force (V) and bending mo-
loading γxz = 0.1 (scale factor: 10). ment (M).
3 Using periodic boundary conditions
The use of PBC allows warping of the cross section, which reduces or even avoids boundary eects
in the RVE. In a two-scale shell model, the introduced displacement for a shear deformation along with
permissible periodic displacements result in rigid body movements. Therefore, shear stiness components
and associated stresses are not obtained, Fig. 1 (PBC).
Due to the previously established modication of the

matrix AI , there is no rotation of the cross section α=0
in the RVE center. The rigid body movement can be
prevented by another constraint (C2 in diagrams)
for the in-plane normal stress. This assumes geo- z
metrical and physical linearity. The constraint is
introduced via an additional row of elements with
nodes in the center plane of the RVE, Fig. 4. The
α
constraint aects the system only via these nodes.
Therefore, only for these the stresses σ̄α are deter-
mined by means of a modied kinematics, which uses ψα , ωα
a common rotation respectively translation of that
plane [4]. These elements do not introduce any fur- Figure 4. Additional elements to apply constraint C2.
ther stiness to the system
Z
(σα − σ̄α ) · [zψα + ωα ] dV = 0 with α ∈ {x, y}. (10)
Ṽα
This constraint has to be established for both in-plane directions. The constraint only needs to be
fullled on average. This still allows warping of the cross section and, thus, the algorithm obtains not
only correct membrane and bending stinesses but also correct shear stinesses, Fig. 1 (PBC+C1+C2).
The constraints C2 leads to four additional degrees of freedom.
Material Geometry Material Geometry

E1 = 2222 N/mm 2
ν1 = 0.1 h1 = 1 cm E1 = 3300 N/mm 2
ν1 = 0.2 h1 = 2 cm
E2 = 100 N/mm 2
ν2 = 0.3 h2 = 3 cm E2 = 1100 N/mm 2
ν2 = 0.3 h2 = 12 cm
E3 = 2222 N/mm 2
ν3 = 0.1 h3 = 1 cm E3 = 3300 N/mm 2
ν3 = 0.2 h3 = 2 cm
Figure 5. Material and geometrical properties for Figure 6. Material and geometrical properties for
the three-layered RVE. the cylindrical panel.
439
5
[−]
DBC DBC: PBC+C1+C2:
/D11,Ref. 4 DBC+C1 12.0
3 PBC
PBC+C1+C2
γ
2 τxz
D11,i
1 [N/cm2 ]
γ
0 0.0
012 4 8 16 32
RVE length to height [−] Figure 8. Shear stress for γxz = 0.1 (scale factor: 2).
Figure 7. Relative shear stiness of a three-layered RVE.
The rst example deals with a three-layered RVE. The two outer, stier layers enclose a soft core, Fig.
5. An investigation of the shear stiness depending on the in-plane sizes is shown in Fig. 7. It shows
again, that the shear stiness for DBC decreases from a too sti behavior to zero. Using constraint C1,
the shear stiness converges to the correct solution, although very slowly. Only the newly developed
approach leads to the correct shear stiness Dγ11 = 209.6 N/cm immediately.
The inuence of the boundary conditions is also evident when considering the shear stress distribution,
as shown in Fig. 8. The distributions are totally dierent which can be of crucial importance considering
local eects.
Within the second example, a cylindrical panel, Fig. 9, with a three-layered structure, Fig. 5, is subjected
to a concentrated load using the arc length method, Fig. 10 . The reference solutions are computed using
the model in [2].
F, w
4
3
Load F [kN]
254
254 2540 2
1 Ref.
θ θ PBC+C1+C2
0
1 θ = 0.1 rad
2 0 5 10 15 20 25 30
[mm] 3
Displacement w [mm]
Figure 9. Cylindrical panel subjected to a concen- Figure 10. Load-displacement-diagram for the
trated load. cylindrical panel.
Conclusion
In this contribution a coupled two-scale shell model with an application of the FE2 -method is presented.
The introduced additional constraints prevent a length dependency of the shear results for DBC and
enable the use of PBC. In addition to the correct membrane and bending results, this allows a correct
calculation of the shear results, as shown in the examples.
References
[1] F. Gruttmann and W. Wagner. A coupled two-scale shell model with applications to layered struc-
tures. International Journal for Numerical Methods in Engineering 94 (2013), pp. 12331254.
[2] F. Gruttmann, G. Knust, and W. Wagner. Theory and numerics of layered shells with variation-
ally embedded interlaminar stresses. Computer Methods in Applied Mechanics and Engineering 326
(2017), pp. 713738.
[3] S. C. Klarmann. Geometrisch und physikalisch nichtlineare Mehrskalenmodellierung räumlicher Stab-
tragwerke. In: Forschungsberichte des Instituts für Mechanik der Technischen Universität Darmstadt.
Bd. 44. Darmstadt: Technische Universität Darmstadt, 2018, pp. 1163.
[4] S. C. Klarmann and F. Gruttmann. A homogenization approach for beam-like structures with arbi-
trarily shaped and deformable cross-sections. Proceedings of the 7th GACM Colloquium on Compu-
tational Mechanics (2017), pp. 476479.
440
MS 22: Recent Trends in Phase–Field Modeling: Theory, Numerics, Applications

An efficient matrix-free finite element solver

for the Cahn–Hilliard equation
1
G.F. Castelli ∗
and W. Dörfler1
1 Karlsruhe Institute of Technology (KIT), Institute of Applied and Numerical Mathematics 2, Karlsruhe, Germany
Abstract.
We focus on the efficient numerical solution of the mixed formulation of the Cahn–Hilliard equation. The
main challenges arise from the handling of an almost sharp moving phase boundary and processes on different
time scales. We employ a higher order finite element method together with an adaptive time integrator. In
particular, we use a fully parallelizable highly efficient matrix-free framework, which avoids the necessity of
storing matrices.
1 Introduction
In the last few years the phase-field method has become a popular approach for modeling various types
of problems with a moving phase boundary. One of the main advantages is that a sharp interface gets
replaced by a smooth transition zone and thus front-tracking methods become unnecessary [5]. To name
just a few examples, the phase-field method was already used to model microstructure evolution, phase
separation, dendritic growth, growth of cancerous tumors and fracture, see [5] and the references cited
therein.
We consider the phenomenon of phase separation in electrode particles of lithium ion batteries during
the (dis-)charge process, because due to changes in the host material, for example LiFePO4 , a separation
into lithium rich and lithium poor phases is possible. This behavior can be described with a phase-field
model resulting in the Cahn–Hilliard equation, see [6, 9]. In a previous work we also considered phase
transition in electrode materials of lithium ion batteries with a sharp interface model and developed a
finite element solver for the coupled electrolyte-electrode system [2].
2 Model Equations
We briefly discuss the model equations introduced in [6], where the temporal evolution of the lithium
concentration including the phase separation and diffusion with two coexisting phases was described with
the Cahn–Hilliard equation.
Let Ω ⊂ R3 be the domain of one electrode particle. In terms of the normalized lithium concentra-
tion c : R≥0 × Ω → [0, 1] the conservation of mass yields:

∂t c − ∇· M (c)∇µ = 0 in R>0 × Ω,
where M (c) = D0 c(1 − c) is the concentration-dependent mobility with the diffusion coefficient D0 > 0
and µ = δΨ/δc is the chemical potential, which is defined as the variational derivative of the system’s
free energy:
Z
Ψ(c, ∇c) = ψ ch (c) + ψ gd (∇c) dx.
Ω
∗ Corresponding author: fabian.castelli@kit.edu
441
Here the homogeneous chemical free energy density is represented by a potential of doublewell-type
1
ψ ch (c) = α1 c + α2 c2 + c log(c) + (1 − c) log(1 − c),
2
and depends on the two parameters α1 and α2 , which control the shape of the doublewell. The gradient
energy density, responsible in the end for the fourth-order derivative, is given by
1
ψ gd (∇c) = κ|∇c|2 ,
2
where the interfacial energy coefficient κ > 0 controls the width of the phase boundary.
The problem reads in its mixed formulation: Find the concentration and the chemical potential satisfying:

∂t c = ∇· M (c)∇µ in R>0 × Ω,
µ = ∂c ψ ch (c) − κ∆c in R>0 × Ω,
M (c)∇µ · n = µext on R>0 × ∂Ω,
∇c · n = 0 on R>0 × ∂Ω,
c(0, · ) = c0 in Ω.
Note, that the lithium insertion, or extraction, is modeled by an inhomogeneous Neumann-type boundary
condition. At the initial state we assume a given concentration profile c0 : Ω → [0, 1], consistent with the
boundary conditions.
3 Numerical Method
In the following we restrict ourself to the classical setting of the Cahn–Hilliard equation, i. e., assuming
homogeneous Neumann-type boundary conditions:
M (c)∇µ · n = ∇c · n = 0 on R>0 × ∂Ω,
and describe the developed finite element solver for this case. Among others, difficulties for the numerical
solution are the higher order differential operator as well as the almost sharp moving phase boundary
and the different time scales.
In space, we use a Lagrangian finite element method of higher order for both, the concentration and
the chemical potential. Hence for curved boundaries of the triangulated domain we also have to use a
correspondingly higher order boundary approximation to get the full convergence order. One advantage
of the proposed mixed formulation is that we avoid the fourth-order differential operator ∆2 and hence
we can circumvent the necessity of C 1 -conforming methods, see [9].
For the Cahn–Hilliard equation using an appropriate time integration scheme is very important, since
several processes take place on different time scales. For example the separation of the phases and the
evolution of the phase boundary happen quasi-instantaneous, while the clustering and coarsening due to
the minimization of the surface free energy can take a long time. A well known time stepping scheme
for the Cahn–Hilliard equation was derived by Eyre [4] and is based on a splitting of the homogeneous
chemical free energy density into a convex and a non-convex part. This scheme is first-order accurate and
unconditionally stable. However, we observed, that we need a very small time step to capture the correct
evolution of the phase boundary. So instead we use the adaptive version of the second-order accurate
unconditionally stable time integrator proposed by [5].
The fully discrete system for one time step with the implicit Euler scheme reads then as follows: Given the
concentration ck and the chemical potential µk at the previous time step tk , find the concentration ck+1
and the chemical potential µk+1 for the next time step tk+1 = tk + τ such that the following weak
formulation is fulfilled:
k+1
vh , ck+1
h − c k
h + τ ∇v h , M c k+1
h ∇µh =0 for all vh ∈ Vh ,

wh , µk+1
h − ∂c ψ ch ck+1
h − κ ∇wh , ∇ck+1h =0 for all wh ∈ Vh .
To this nonlinear problem we apply the Newton method. For small problems we can use direct solvers
such as the LU-decomposition for the computation of the Newton updates. However, a full resolution
of the three-dimensional spatial domain, respecting the almost sharp phase boundary, generates a huge
amount of degrees of freedom.
442
3.1 Matrix-free Algorithms
For the numerical linear algebra we use the highly efficient matrix-free implementation, provided by [7],
within of the open-source finite element library deal.II [1]. To briefly explain the idea of matrix-free
algorithms we consider for a moment a generic time-dependent partial differential equation. If we use an
implicit time stepping scheme, we will end up with the generic problem to solve a linear system A x = b in
almost every case. If we think of iterative methods such as the CG or the GMRES method, we know that
these are based on Krylov subspaces and need matrix-vector multiplications only [8]. Now the key idea is
to implement a function evaluating the matrix-vector product without storing the matrix itself. Although
this makes the recomputation of the matrix entries unavoidable, every time a matrix-vector product is
evaluated, this is much faster on todays modern processor architectures, since the limiting factor for
matrix-vector operations is mostly not the arithmetic throughput, but the memory bandwidth [7].
To show the performance of the matrix-free implementation we solved a nonlinear test problem:
−∆u = 3 exp(u) in Ω = (0, 1)2 ,

u=0 on ∂Ω,
which is in some sense representative for the solution of one time step of the Cahn–Hilliard equation. In
Figure 1 we plotted the computational times against the number of cores for different problem sizes. As
this figure shows, we get an optimal strong scaling behavior of the parallelization.
102
4.2M DoFs
16.8M DoFs
67.1M DoFs
Linear scaling
Time solver [s]
101
100
10−1 0
10 101 102
Number of cores
Figure 1. Strong scaling of the parallel computation.
However, since we do not have access to matrix entries, we can only use iterative Krylov subspace
methods to solve the linear systems. Furthermore there is a high demand for good preconditioners, which
are matrix-free implementable.
4 Numerical Results
To demonstrate the capabilities of our developed matrix-free finite element solver for the Cahn–Hilliard
equation, we apply it to a test problem from the literature, Problem 4 of [3]. As domain we have Ω =
(0, 1)2 the two-dimensional unit square and let the final time of the simulation be T = 10. The homoge-
neous chemical free energy density is replaced by a quartic polynomial ψ ch (c) = 1/4 c2 (1 − c)2 and the
mobility by a constant M = 0.25. The interfacial energy coefficient is set to κ = 10−4 . As in Section 3
we use homogeneous Neumann-type boundary conditions and as initial condition a smooth perturbation
of 0.5. We triangulate the domain with 16k uniform rectangles and use fourth-order Lagrangian finite
elements, i. e., in total we distribute 526k degrees of freedom. With 16 cores of a parallel cluster we are
able to solve this problem with 330 adaptive time steps in less the four hours. The initial and final state
of the solution are depicted in Figure 2. The fast solution of the Cahn–Hilliard equation is also due to our
new developed preconditioner, with which we need also for this large problem at maximum ∼ 70 GMRES
iterations and even much less as the phase separation dynamic relaxes.
5 Conclusion
Summing up, we discussed the equations modeling phase separation in electrode particles of lithium
ion batteries during the (dis-)charge process. Further we presented a numerical method to solve the
443
Figure 2. Concentration profile of the Cahn–Hilliard test problem at initial (left) and final (right) time.
Cahn–Hilliard equation and explained the idea of the used matrix-free framework. Finally we showed
the parallel scaling behavior of the matrix-free implementation and solved a test problem for the Cahn–
Hilliard equation.
With this efficient finite element solver we are now able to simulate lithium insertion into full three-
dimensional electrode particles. In particular, we can now allow a tensor-valued mobility and thus
investigate also anisotropic diffusion. Further we can allow general non-uniform insertion rates. In
such a case, with anisotropic mobility and non-uniform insertion rate, the formation and the migration
of the phase boundary through an electrode particle will be very interesting. Another advantage of our
implementation is the geometry independence, this means we can also simulate arbitrary shaped particles.
To further reduce the computational time and increase the efficiency we want to incorporate spatial
adaptivity based on residual-type a posteriori error estimators. In collaboration with colleagues from
SiMET we currently include a new thermodynamic consistent nonlinear solid mechanics theory taking
swelling and stress generation for large strains into account, however, this is part of a future work.
Acknowledgments
The authors thank M. Kamlah and T. Zhang for discussions about the model equations. G. F. Castelli
acknowledges financial support by the German Research Foundation (DFG) through RTG 2218 SiMET –
Simulation of mechano-electro-thermal processes in lithium-ion batteries, project number 281041241.
References
[1] W. Bangerth, R. Hartmann, and G. Kanschat. deal.II—a general-purpose object-oriented finite ele-
ment library. ACM Trans. Math. Software 33.4 (2007), pp. 24/1–24/27.
[2] G. F. Castelli and W. Dörfler. The numerical study of a microscale model for lithium-ion batteries.
Comput. Math. Appl. 77.6 (2019), pp. 1527–1540.
[3] M. Dehghan and V. Mohammadi. The numerical solution of Cahn–Hilliard (CH) equation in one, two
and three-dimensions via globally radial basis functions (GRBFs) and RBFs-differential quadrature
(RBFs-DQ) methods. Eng. Anal. Bound. Elem. 51 (2015), pp. 74–100.
[4] D. J. Eyre. Unconditionally gradient stable time marching the Cahn–Hilliard equation. MRS Pro-
ceedings 529 (1998), pp. 39–46.
[5] H. Gomez and T. J. R. Hughes. Provably unconditionally stable, second-order time-accurate, mixed
variational methods for phase-field models. J. Comput. Phys. 230.13 (2011), pp. 5310–5327.
[6] M. Huttin and M. Kamlah. Phase-field modeling of stress generation in electrode particles of lithium
ion batteries. Appl. Phys. Lett. 101.13 (2012), pp. 133902/1–133902/4.
[7] M. Kronbichler and K. Kormann. A generic interface for parallel cell-based finite element operator
application. Comput. & Fluids 63 (2012), pp. 135–147.
[8] Y. Saad. Iterative Methods for Sparse Linear Systems. 2nd ed. SIAM, Philadelphia, PA, 2003.
[9] Y. Zhao, P. Stein, and B.-X. Xu. Isogeometric analysis of mechanically coupled Cahn–Hilliard phase
segregation in hyperelastic electrodes of Li-ion batteries. Comput. Methods Appl. Mech. Engrg. 297
(2015), pp. 325–347.
444

Improving the numerical solution of the phase-field equa-

tion by the systematic integration of analytic properties
of the phase-field profile function
1
Michael Fleck ∗
, Felix Schleifer1 , Markus Holzinger1 and Uwe Glatzel1
1 Metals and Alloys, University of Bayreuth, Germany
Abstract. Different Allen-Cahn type phase-field models are investigated within a unified finite difference
framework. We consider the stationary growth of arbitrarily oriented planar interfaces in 3D driven by a
constant chemical potential. The different models are classified with respect to their individual ability to
accurately reproduce the analytically exact growth behavior, with regard to the transformation velocity, the
phase-field width as well as the interface energy density. We find superior behavior and accuracy for the
models that have a computation-grid-dependent equilibrium potential providing translational invariance of
the phase-field profile in the direction of growth, as recently suggested by A. Finel and co-workers [3]. Base on
that, we propose a new phase-field model, which provides the translational invariance in the direction normal
to the interface based on a local measure of the interface orientation. The new model shows the expected
superior properties arbitrarily oriented planar interfaces also in the cases where the interface is not aligned
with the principle directions of the numerical lattice.
Introduction
The phase-field method offers a very versatile way to setup models for the simulation of microstructure-
evolution involving interfaces, that may undergo topological changes. The key point providing the ver-
satility is the diffuse interface, where the phase-field variable varies smoothly, for instance, from zero
to one about a specific length called the phase-field width. However, the rapid spatial variation of the
phase-field solution naturally leads to difficulties in the numerical computation [1, 5].
Considering the stationary growth of a planar interface driven by a constant chemical potential, we quan-
titatively investigate the deviations of the numerical solution from the respective analytic expectations.
Especially, when going over to sharper and sharper interfaces, we observe increasingly large deviations up
to the extrem situation of a vanishing interface velocity that corresponds to the situation of grid pinning.
Recently, A. Finel and co-workers proposed a new way to eliminate the related spurious grid-adhesion
forces in 1D resulting from the spatial discretization on the numerical lattice [3].
The method of how a translationally invariant phase-field formulation in 1D is setup is briefly reviewed in
section 1. Further, we quantitatively compare the growth behavior of a constantly driven planar interface
using the new 1D model with respective results obtained from a conventional phase-field formulation. We
find, that this new class of phase-field models provides superior properties and accuracy, even in the case
of very sharp interfaces. In section 2 we discuss the stationary growth of planar interfaces with arbitrary
orientation. We propose a new phase-field model of this class, which provides the translational invariance
in the growth direction, based on the local measure of the interface orientation. Especially for the case
of the very sharp interfaces, we also propose a new and more accurate measure of the interface position
based on informations of the nonlinear phase-field profie, as discussed in section 3. Finally, in section
4, we discuss the special phase-field boundary conditions, where the interface is forced to take a specific
wetting angle with the respective boundary. These boundary conditions are required for the realization
of the configuration of a planar interface growing into arbitrary directions.
∗ Corresponding author: michael.fleck@uni-bayreuth.de ORCID iD: 0000-0002-2799-9075
445
1 Elimination of grid adhesion in 1D

In the simplest case, we may consider the transformation between just two phases, discriminated by a
single phase-field. The starting point is a volume integral over an energy density, which locally depends
on the phase-field value and gradient. The energy density is additively decomposed into a phase-field
gradient contribution, an equilibrium potential contribution and a local non-equilibrium driving force
contribution.
On a discrete numerical lattice the phase-field is represented by a finite array of values that are distributed
on grid points within the computational lattice. We consider a planar interface in 3D with a hyperbolic
tangent like profile of the following form

1 2 (i∆x − x0 )
φi = 1 + tanh , (1)
2 ξ
where i denotes the discrete locations of the grid points within the numeric lattice. The central analytic
property of the hyperbolic tangent profile (1), can be easily derived form the addition theorem of the
hyperbolic tangent function [3, 4]:
(2φi − 1) ± a
2φi±1 − 1 = , (2)
1 ± (2φi − 1) a
where the grid-coupling parameter a is defined as

2∆x
a = tanh . (3)
ξ
In 1D static interfaces, such as Eq. (1), have to satisfy the discrete equilibrium condition as given by
ξ2 ∂g(φi )
(φi+1 − 2φi + φi−1 ) = . (4)
∆x2 ∂φ
The key point is to identify a double-well potential g(φ) such that, if φi = f (i∆x) is solution of Eq. (4),
then φi = f (i∆x − x0 ) is also a solution for any real translation x0 [3]. Then, the phase-field profile
will be obviously invariant under a translation providing an interface free of any spurious grid-adhesion.
Interestingly, the analytic property Eq. (2) provides the following modification of the derivative of the
equilibrium potential [3, 4]
∂g(φ) a2 ξ 2 4φ (1 − φ) (1 − 2φ)
= , (5)
∂φ ∆x2 1 − a2 (1 − 2φ)2
such that the equilibrium condition (4) is satisfied for arbitrary choices of the interface position x0 . Note,
that in the continuum limit ∆x → 0, this potential converges to the original derivative of the quartic
double well potential, ∂g/∂φ = 16φ (1 − φ) (1 − 2φ), as lim∆x→0 ξ 2 a2 /∆x2 = 4.
1 1 1
TI model 0.1 TI model new model
velocity error: 1 - v/vth
Energy error: 1 - Γ/Γth
0.1 0.1
Width error: 1 - ξfit/ξ
conventional conventional conventional

0.01
0.01 0.001 0.01
0.001 0.0001 0.001
0.0001 1x10-5 0.0001
1x10-6
1x10-5 1x10-7 1x10-5
1x10-6 1x10-8 1x10-6
0 2 4 6 8 10 12 0 2 4 6 8 10 12 0 2 4 6 8 10 12
a) phase-field width: ξ/Δx b) phase-field width: ξ/Δx c) phase-field width: ξ/Δx
Figure 1. Comparison of the modeling behavior with and without translational-invariance during the simulation
of stationary growth of the planar front in 1D.
Considering the stationary growth of the planar interface driven by a constant chemical potential, we
show that the modified derivative of the equilibrium potential Eq. (5) indeed eliminates all spurious
grid-adhesion forces on interfaces as represented by the profile (1). It turns out that the elimination
of spurious grid-adhesion forces allows to chose the width of the diffusion interface region as small as
ξ/∆x = 0.6, i.e. nearly half of the grid spacing. In Fig. 1, we compare the deviations of the resulting
transformation velocity, the interface energy as well as the fitted phase-field width from the respective
446
analytic expectations for the conventional phase-field model with the result from the new model with
translational invariance within a unique finite difference framework. It is quite surprising that, even for
values of the phase-field width as small as ξ/∆x = 0.6 the error of the results from the translationally
invariant model is still orders of magnitude lower than that of the conventional model for ten-times wider
interfaces.
This method for the elimination of the grid-adhesion forces in the phase-field equation is related to
earlier suggestions of Glasner and Weiser. Bother authors proposed the idea to improve the numerical
performance of phase-field solvers via a nonlinear scaling of the phase-field equation by the profile function
[5, 6]. Bother authors report about highly improved properties and accuracies of their solvers, after the
nonlinear transformation using the nonlinear analytic profile function.
We point out that also for other profile-functions such as the section-wise defined sinus-profile such a
modification of the equilibrium potential to diminish spurious grid-adhesion is feasible. Interestingly, we
find that a respective derivation for the section-wise defined sinus profile does not provide a new function,
but provides again the section-wise quadratic double obstacle potential, now with a computation-grid-
dependent correction factor, as had been previously proposed by Eiken [2].
2 Arbitrarily oriented planar interfaces

1 10
conventional
0.01 1
velocity error: 1 - v/vth
amplitude in a.u.
fixed TI model
Pinning forces
0.0001 0.1 new model

1x10-6 0.01
1x10-8 0.001
1x10-10 fixed TI model 0.0001
1x10-12 new model 1x10-5
1x10-14
1x10-6
0 5 10 15 20 25 30 35 40 45 0 2 4 6 8 10 12
a) interface orientation θ in ° b) phase-field width: ξ/Δx
Figure 2. a) Plot of the fluctuation amplitude of the total grid-adhesion forces as function of the interface ori-
entation angle for ξ/∆x = 1 and b) Comparison of the relative error in the resulting transformation
velocity for an interface growing under an angle of ϑ = 45◦ with respect to the basic lattice vectors of
the computational grid.
Next, we investigate the total grid-adhesion forces acting on arbitrarily oriented planar interfaces in 3D.
We find that the translational invariance as well as all the related remarkable improvements provided by
the modified equilibrium potential Eq. (5) is tight to an extremely sharp window of interface orientations
around the initial direction, for which the translational invariance has been restored. The sharpness
of this orientation window can be seen in Fig. 2a), where the solid purple curve shows the fluctuation
amplitude of the total grid-adhesion forces as function of the interface orientation, when the direction
of translational invariance is kept as fixed. These total grid-adhesion forces correspond to an integral
over the right hand site of the phase-field equation, measured for a single interface with an ideal analytic
profile. When the discrete representation of the analytic profile Eq. (1) is moved along the interface
normal direction, such that its center x0 is passing several grid points, the measured total grid-adhesion
forces oscillate with the mentioned fluctuation amplitude.
In order to overcome the problems with a too sharp orientation window, we propose a new model, which
measures the local interface orientation first, and then restores translational invariance for this locally
measured direction. For this purpose the interface orientation has to be measures quite accurately, even
for very small choices of the phase-field widths. Therefore, again we make use of the addition property
−
Eq. (2), for the local calculation of the ak . Moreover, we average ak = a+k + a k /2, where
± (φi±uk − φi )
a±
k = ,
φi − 2φi±uk φi + φi±uk
where the lattice vector uk connects two neighboring grid points in a certain direction within the numerical
lattice. ∆xk = |uk | is the distance between two neighboring grid points in the respective lattice-direction.
We point out, that the orientation of the interface is uniquely described by the unit normal vector n. This
leads to respective constrains on the grid-coupling parameters ak , that have to be satisfied. Therefore, we
calculate the k−projections of the normal vector, with nk = ξ arctanh ak /(2∆xk ), enforce the unit length
of the orientation vector and re-calculate the new grid-coupling parameters ak based on the corrected
interface normal vector. The green curve in Fig. 2a) shows that for the respectively new formulation the
fluctuation amplitude of the grid-adhesion forces is indeed significantly reduced. Furthermore, in Fig. 2b)
447
we show that the relative error on the stationary phase transformation velocity for a planar front growing
under an angel of ϑ = 45 deg is equally small as compared to the 1D-case shown in Fig. 1a), which proves
that the new model provides grid-adhesion free motion of planar interfaces in 3D.
3 Measure of the interface position

For the results in Fig. 1 as well as Fig. 2 b), it should be mentioned that the conventional measure
of the interface position by a linear interpolation of the phase-field values at neighboring grid points is
not sufficiently accurate. Especially, when the phase-field width becomes very small, we have to non-
linearly interpolate the interface position using the phase-field profile function 1. We propose to calculate
interpolated position, where the phase-field profile takes the arbitrary contour-level l (0 < l < 1), as

int λ l − φi
,
xi = i∆x + arctanh (6)
2nk 2lφi − φi − l
where nk denotes the projection of the unit normal interface vector onto the k−th numerical lattice-
direction, between which the contour position has to be calculated. In order to further regularize the
resulting interface position as function of time, we impose a linear interpolation between the two slightly
different interface positions xint
i and xint
i+uk of the two neighboring grid points, connected by the lattice
vector uk , next to the position of the contour-level l . In this case, we have the condition (φi − l) ·
(φi+uk − l) ≤ 0.
4 Wetting angle boundary conditions for the phase-field

In order to realize stationary growth of the planar front with arbitrarily oriented interfaces, that are
explicitly not aligned with basic lattice-vectors of our the simple cubic numerical lattice, we propose new
a phase-field boundary condition, which enforces a certain wetting angle ϑ with the respective boundary.
For a boundary plane with orientation k the phase-field value at a boundary grid point i + uk has to
calculated from the respective value at the neighboring field grid point i. The profile-shift by the length
sn = ∆xk sin ϑ along the interface normal direction can be realized using the addition theorem for the
hyperbolic tangent Eq. (2).
(1 − aϑ ) φi
φi+uk = , (7)
1 − aϑ (2φi − 1)
where the grid-coupling parameter is now defined as aϑ = tanh (2∆xk sin ϑ/λ).
Acknowledgments
We thank the Federal Ministry for Economics and Energy (BMWi) of the Federal Republic of Germany
for the financial support under the running project COORETEC: ISar (funding code: 03ET7047D).
Further, we thank the Federal Ministry of Education and Research (BMBF) for the financial support
under the running project ParaPhase (funding code: 01IH15005B). Also, the financial support from the
German Research Foundation (DFG) within the second phase of the priority program 1713 (FL 826/3-1
| GL 181/53-1) is greatly acknowledged. Then, we are very thankful to our collaboration partners at
ACCESS, especially B. Böttger and J. Eiken, MTU Aero Engines AG and Siemens AG. Finally, we
would like to thank A. Finel from ONERA, CNRS, University of Paris-Saclay in France for the vivid and
fruit-full discussions on this issue.
References
[1] A. Bösch, H. Müller-Krumbhaar, and O. Shochet. Phase-field models for moving boundary prob-
lems: Controlling metastability and anisotropy. Z. Phys. B 97 (1995), pp. 367–377. doi: 10.1007/
BF01307490. url: https://doi.org/10.1007/BF01307490.
[2] J. Eiken. Numerical solution of the phase-field equation with minimized discretization error. IOP
Conf. Series: Materials Science and Engineering 33 (2012), p. 012105. doi: doi:10.1088/1757-
899X/33/1/012105.
[3] A. Finel, Y. Le Bouar, B. Dabas, B. Appolaire, Y. Yamada, and T. Mohri. Sharp Phase Field
Method. Phys. Rev. Lett. 121 (2018), p. 025501. doi: 10.1103/PhysRevLett.121.025501. url:
https://link.aps.org/doi/10.1103/PhysRevLett.121.025501.
448
[4] M. Fleck, F. Schleifer, and U. Glatzel. Diminishing of numerical grid-adhesion in the phase-field
equation for marginally diffuse interfaces. Submitted to PRL (2019).
[5] K. Glasner. Nonlinear Preconditioning for Diffuse Interfaces. J. Comp. Phys. 174 (2001), pp. 695–
711. doi: https://doi.org/10.1006/jcph.2001.6933. url: http://www.sciencedirect.com/
science/article/pii/S0021999101969335.
[6] M. Weiser. Pointwise nonlinear scaling for reaction–diffusion equations. Appl. Num. Math. 59 (2009),
pp. 1858–1869. doi: https : / / doi . org / 10 . 1016 / j . apnum . 2009 . 01 . 010. url: http : / / www .
sciencedirect.com/science/article/pii/S0168927409000129.
449
450

Phase-field damage model coupled with ageing in

polymer adhesives due to moisture diffusion
∗
Siva Pavan Josyula and Stefan Diebels
Applied Mechanics, Saarland University
Abstract. This paper deals with the formulation and numerical implementation of a coupled continuum model
for deformation, diffusion and damage in viscoelastic materials like polymer adhesives to analyse the influence
of ageing behaviour on damage phenomena. The formulation is based on a continuum mechanical model.
In order to couple diffusive moisture transport with viscoelasticity, a Langmuir-type diffusion model is used,
since the diffused moisture in polymers exists in two phases namely free and bound moisture. Furthermore,
the coupled deformation-diffusion model takes into account the decrease in stiffness and relaxation times with
increasing moisture content. Finally, the model is coupled with damage using a phase-field approach. The
governing equations are obtained from supplementing balance equations with thermodynamically consistent
constitutive theory, where displacements, phase-field order parameter and moisture concentration are the
primary variables. Finally, degradation of the material is chosen as a fully coupled function of the moisture
concentration. The numerical implementation is based on the finite element method in space and an implicit
scheme for the temporal discretization. Some examples show the applicability of the presented model.
Introduction
The applications of the polymers in the commercial industry are rising especially in the manufacturing
of lightweight structures. For technical applications, it is necessary to develop a computationally reliable
material model to analyze the mechanical behaviour. Early ageing in the polymer adhesives showcases
an important role in their application for commercial purpose, as it is apparent to lose their cohesive
properties and degrades earlier than the expected lifetime.
In the current study, a polymer is considered to be exposed to humid environmental conditions at a
constant temperature. These environmental conditions reflect their influences on the ageing of polymers
by affecting mechanical properties, demonstrating both physical and chemical ageing. Physical ageing
is due to moisture diffusion, which is reversible in the drying process. Unlike physical ageing, chemical
ageing is an irreversible process caused due to breakage of the existing bonds or by the creation of new
bonds. The degradation of physical ageing in polymers is modelled with a phase-field damage model.
1 Numerical formulation of a phase-field damage model for ageing in polymers
Here, we discuss the basic notations and the formulation for a phase-field damage model coupled with
moisture concentration and mechanical properties [2].
The degradation of the material is described by the phase-field variable 0 ≤ φ(x) ≤ 1, where φ(x) = 1
describes the damage and φ(x) = 0 fully intact material, respectively.
The total potential energy function, is given by
Z Z Z
Ψpot = Ψm (ε, φ, c) + Ψs (φ, c) + Ψc (c) (1)
Ω Ω Ω
where the first term Ψm is the stored bulk energy, the second term Ψs refers to the surface energy asso-
ciated with the formation of the crack and the last term Ψc refers to the relative moisture concentration.
∗ Corresponding author: siva.josyula@uni-saarland.de
451
Due to the degradation of mechanical behaviour, the material experiences the loss in the stiffness, which
is characterized by the degradation function g(φ).
g(φ) = (1 − κ)φ2 + κ (2)
A simple quadratic function in φ is chosen as a degradation function such that it decreases monotonically.
The degradation function g(1) = 1 represents the intact and g(0) = 0 represents fully damaged material,
respectively. Furthermore, the condition g 0 (0) = 0 controls the contribution of the bulk energy in the
phase-field evolution. In Equation 2 a regularization parameter κ is introduced for a well-posed stable
model in case of the fully degraded material.
1.1 Moisture transport and Diffusion
The diffusion-driven moisture absorption in PU shows anomalous behaviour, and the diffused moisture
exists in two phases characterized as bound and free moisture. The moisture present around the polymer
network making a strong physical bond, is characterized as bound moisture. Unlike the bound moisture,
the free moisture is far from the polymer network and free to diffuse.
The Ficks’s second law of diffusion falls in-short to consider trapped moisture in phases. Therefore, we
consider a non-Fickian diffusion model known as a Langmuir-type model. In this diffusion model, the
total moisture concentration is split into free and bound moisture c = cf + cb
∂cf ∂cb
+ = D∇2 cf (3)
∂t ∂t
Here, the evolution equation of the bound water concentration can be formulated as
∂cb
= γcf − βcb (4)
∂t
where γ refers to the rate at which the mobile moisture becomes bound, and β refers to the rate at which
the bound molecules of moisture become mobile.
Since, we consider that the moisture content present in the environment is free and sufficient in quantity,
the value of the Dirichlet boundary condition cf (ΓD , t) follows cb in each time step from the difference
between the total moisture Nmax and the quantity of bound water at the boundary ΓD
cf (ΓD , t) = Nmax − cb (ΓD , t) (5)
1.2 Mechanical behaviour of polymer
A material model based on the continuum approach is used to describe viscoelastic behaviour. In the
simplest case, a 3-parameter rheological model can describe the viscoelastic behaviour consisting of springs
and a damper, as shown in Figure 1. The moisture dependent stiffness parameters are used to describe
stiffness in springs by a generalized Hookes’s law [4] and the stress in the damper is governed by the
strain rate.
Figure 1. 3-Parameter model of viscoelasticity
452
A sigmoidal function f (c) ∈ [0 1] is used to compute the stiffness paramters ndependent of ageing
numerically. Calculation of the stiffness paramters is shown in Equation 6.
λ(c) = f (c)λdry − (1 − f (c))λsat

µ(c) = f (c)µdry − (1 − f (c))µsat
k1 (6)
f (c) = 1 −
1 + exp(−k2 (c − k3 ))
where c = cf + cb and k1 , k2 , k3 are constants
The model reads as,
ε = εe + εi σ = σ eq + σ neq
σ eq = 2µ (c)ε + λ0 (c)trεI
0 (7)
σ neq = 2µ1 (c)(ε − εi ) + λ1 (c)tr(ε − εi )I.
The viscoelastic material properties used are listed in Tabel 1.
Table 1. Viscoelastic properties of polymer adhesive
dry 50 % r.H saturated

λ0
5.3 MPa 2.82 MPa 1.2 MPa
equilibrium
µ0 3.38 MPa 1.79 MPa 0.76 MPa
λ1 13.5 MPa 8.78 MPa 0 MPa
nonequilibrium
µ1 8.6 MPa 5.59 MPa 0 MPa
1.2.1 Hygroscopic swelling due to moisture diffusion
Due to the hygroscopic phenomenon in polymers, they swell due to an increase in the total volume
during the absorption process.. These changes in the material due to hygroscopic swelling results in
strains known as the swelling strains.
A linear function dependent on the swelling strains and relative concentration is defined to characterize
the swelling behaviour [3].
εswell = αc (8)
where c = cf + cb is the total moisture concentration trapped inside the material and α is the constant
known as coefficient of hygroscopic swelling.
1.3 Surface energy degradation under influence of moisture
The moisture diffused into the material reduce the bonding energy by interacting with the molecules
forming the polymer network. This reduction in bonding energy in-turn reduces the fracture resistance.
Here a linearly decreasing function is considered for computing the fracture energy with a change in the
total concentration of moisture numerically. This function is defined from the critical energy release rate
and the relative concentration [1], as shown in Equation 9.
Gc (c)
= 1 − χc (9)
Gc (0)
χ is an additional constant describing the influence of the moisture on the degradation process.
453
2 Numerical example: Cracked square plate subjected to tension in a humid environment
As an example, we consider a single-edge notched specimen of unit length with a horizontal crack starting
from the left-hand side to the centre of the specimen. The specimen’s bottom edge is constrained in all
direction and a uniformly distributed displacement is applied on the top edge. Finally, the geometry is
exposed to a humid atmosphere from all exterior edges. The applied boundary conditions to simulate
the polymer adhesive at saturated condition is shown in Figure 2a.
(a) Problem setup (b) crack pattern (c) load deflection curve
Figure 2. Damage in test setup under tension
Figure 2b illustrates the damage with blue and red contours. Blue contour corresponds to the damage and
the red contour fully intact state of the material, respectively. Figure 2c show the damage propagation
in the single-edge notched specimen under tension, where it is evident from the force versus displacement
curve that the force reduces after damage.
Acknowledgement
The research project 19730 N “Berechnung des instationären mechanischen Verhaltens von alternden Kle-
bverbindungen unter Einfluss von Wasser auf den Klebstoff” of the research association Forschungsvere-
inigung Stahlanwendung e.V. (FOSTA), Düsseldorf was supported by the Federal Ministry of Economic
Affairs and Energy through the AiF as part of the programme for promoting industrial cooperative re-
search (IGF) on the basis of a decision by the German Bundestag. The project is carried out at Chair of
Applied Mechanics, Saarland university
References
[1] A. Alvaro, I. Thue Jensen, N. Kheradmand, O. M. Lovvik, and V. Olden. Hydrogen embrittlement
in nickel, visited by first principles modeling, cohesive zone simulation and nanomechanical testing.
International Journal of Hydrogen Energy 40 (2015), pp. 16892–16900.
[2] M. Ambati, T. Gerasimov, and L. De Lorenzis. A review on phase-field models of brittle fracture
and a new fast hybrid formulation. Computational Mechanics 55 (2015), pp. 383–405.
[3] X. Fan. Mechanics of Moisture for Polymers: Fundamental Concepts and Model Study. IEEE (2008).
[4] F. Goldschmidt and S. Diebels. Modelling and numerical investigations of the mechanical behavior
of polyurethane under the influence of moisture. Archive of Applied Mechanics 85 (2015), pp. 1035–
1042.
454

An anisotropic phase-field model based on a Represen-

tative Crack Element (RCE)
Dennie Supriatna1 , Johannes Storm1 and Michael Kaliske ∗ 1
1 Institute of Structural Analysis, Technische Universität Dresden, 01062 Dresden, Germany
Abstract. To predict failure phenomena of brittle material is a challenging task. In brittle failure, no apparent
deformations are visible prior to the formation of the crack. The presence of aspects such as stress concentra-
tions, crack-growth instabilities and displacement field discontinuities yields fracture of brittle material harder
to predict in an accurate way. The crack may also be a result of the fracture mode changes or even mixed
of several fracture modes due to loading applications and geometry. Strobl, Seelig [4] and Steinke, Kaliske
[3] have shown, that the standard phase-field approach with spectral decomposition or volumetric-deviatoric
(V-D) split leads to unrealistic predictions for the stress transfer through the crack and they have developed
a more realistic split for the case of isotropic elasticity and small deformations. The model is constructed
from the kinematics of a discrete crack and is further derived to split the crack driving stresses and persistent
stresses. In the current research, a Representative Crack Elements (RCE) concept is developed based on the
kinematics of a discrete crack model. The proposed RCE approach is coupled to the phase-field method by
adopting the variational homogenization concept of Blanco et al. [1]. The consistent kinematics and general
variational framework enable us to extend the RCE into anisotropic material, finite deformation, dissipative
material and coupled problems with other fields such as temperature. To compare the behaviour of the pro-
posed continuum phase-field model with the RCE and a corresponding discrete model, several self consistency
conditions for isotropic and anisotropic materials are employed. We demonstrate the capability of the ap-
proach by presenting several numerical examples, which cover isotropic, anisotropic and coupled problems.
Introduction
A realistic and accurate model of material degradation within the phase-field approach is currently under
development. It has been shown by Strobl and Seelig [4], Schueler [2] and Steinke and Kaliske [3], that
the classical V-D split and spectral decomposition lead to wrong predictions for the force transfer through
the cracked body. They have proposed a more kinematically consistent approach to overcome the issue.
Steinke and Kaliske [3] have proposed a directional decomposition approach, which is constructed from
the kinematics of a discrete crack and is further derived to split the crack driving stresses and persistent
stresses.
Inspired by the work of Steinke and Kaliske [3], a novel approach based on so-called Representative Crack
Elements (RCE) is derived using the kinematics of a discrete crack model. Due to the general variational
framework and consistent kinematics, this RCE allows investigations of complex crack mechanicsms:
re-contacting after cracking, friction, inelasticity, anisotropy, and finite deformations.
1 Concept of Representative Crack Elements (RCE)
The proposed novel concept is based on a fully equivalent re-interpretation of the strain potential energy
as the interpolation of the potential energies of the intact material ψ 0 and of the fully degraded material
ψ c using the degradation function
ψ mech = [1 − g(s)] ψ c + g(s)ψ 0

= ψ c + g(s) ψ 0 − ψ c . (1)
455
1
2 l1
uΓ
l1 discrete
crack with 1
contact 2 l1 n1
n3
l2 l3 l2 l3
n2
a) b)
Figure 1. a) Sketch of the RCE with a discrete crack in local coordinates and b) definition of crack displacements
and deformations.
Based on the clearly defined material states, the additive decomposition of stress and material tangent
reads

S = S c + g(s) S 0 − S c , (2)
c
0 c

C = C + g(s) C − C , (3)
which are composed of the material response without any crack (S 0 , C0 ) and the material response with
a fully developed crack (S c , Cc ). Once, the two unique material states are identified and described, crack
evolution (driving force) is the result of using the novel concept and no longer of an independent model
assumption.
The fully broken material is modelled by the RCE. The following kinematic coupling operators are chosen
u := ū + H̄ · (η − η 0 ) + ũ, (4)
Z
1
ū := u dV , (5)
V Ω
Z
1
H̄ := H dV . (6)
V Ω
Quantities with bar indicate the continuous phase-field model, whereas the rest is associated with the
RCE. The displacement field of the RCE u depends on the displacement ū, the displacement gradient
H̄ of the material point in the continuous model, as well as on the position vector in the RCE η, a
constant vector η 0 and an unknown displacement fluctuation field ũ. The relations of the displacements
and the displacement gradients towards the continuous model are given as their volume average on the
RCE domain Ω in the reference configuration. Applying the concept of energy conservation, the first
Piola-Kirchhoff stress tensor is computed as
Z
1
P̄ = P dV . (7)
V Ω
The variational problem of the RCE regarding the fluctuation field can be solved using various approaches.
For reasons of computational efficiency, an analytical solution for elastic and thermo-elastic material
behaviour is used in the following.
2 Applications
The RCE model is considered as two blocks of intact material, which are separated by a discrete crack. The
crack undergoes crack contact in case of crack closure (see Fig. 1a)). With the assumption of homogeneous
material and the neglect of softening behaviour in the two blocks, piecewise uniform boundary conditions
yield homogeneous stress and strain states in the two blocks for the open and closed crack configurations.
Thus, all outer RCE surfaces and the crack surfaces remain plane and the deformation of the blocks reads
F = F̄ − Γ
3
X
= F̄ − Γi P i , (8)
i=1
456
where F is the deformation gradient and Γi = uΓi /l1 are the crack deformations obtained from the
relative displacements of the crack surfaces uΓ normalised by the RCE length l1 (see Fig. 1b)). The crack
projectors P i are defined as
P 1 = n1 ⊗ n1 , (9)
P 2 = n2 ⊗ n1 , (10)
P 3 = n3 ⊗ n1 (11)
using the basis vectors ni of the local RCE coordinate system.
The unknown crack deformations Γi can be obtained by minimisation of the total potential energy of the
RCE
Z
1
ψ= P : F dV . (12)
2 Ω
Inserting the kinematic coupling Eq. (8) yields the extremal problem

Γi (F̄ ) = arg min ψ̂(F̄ , Γi ) , (13)
Γi ∈R
without constraints on the crack deformations. In case of crack surface contact (Γ1 < 0), the constrained
minimisation problem is to be solved

Γ2,3 (F̄ ) = arg min ψ̂(F̄ , Γ2 , Γ3 ) for Γ1 < 0, (14)
Γ2,3 ∈R Γ1 =0
Γ1 = 0.
The energy potential of the RCE for the case linear anisotropic elasticity reads
1
ψ̂(Ē, Γi ) = V Ē : C : Ē − 2Ē : C : Γsym + Γsym : C : Γsym
2 " #
X3 X3 X3
1 sym sym sym
= V Ē : C : Ē − 2Ē : C : Γi P i + Γi P i : C : Γi P i .
2 i=1 i=1 i=1
In order to solve the unconstrained minimisation problem of Eq. (13)

∂ ψ̂(Ē, Γi ) !
=0 (15)
∂Γi
has to hold, which yields the system of equations
h i
CΓi [Γi ] = r Γi , i ∈ {1, 2, 3} (16)
with
h i
CΓi = P sym
i : C : P sym
j + P sym
j : C : P sym
i , i, j ∈ {1, 2, 3} (17)
Γi sym
r = P i : C : Ē + Ē : C : P sym
i , i ∈ {1, 2, 3} (18)
h i
for the unknown crack deformation Γi . The inversion of the symmetrical 3x3 matrix CΓi results in the
crack deformations from
h i−1
[Γi ] = CΓi r Γi . (19)
For the case that Γ1 < 0 holds, the constrained problem reads
h i
CΓi [Γi ] = r Γi , i ∈ {2, 3} (20)
with
h i
CΓi = P sym
i : C : P sym
j + P sym
j : C : P sym
i , i, j ∈ {2, 3} (21)
Γi sym
r = P i : C : Ē + Ē : C : P sym
i , i ∈ {2, 3} (22)
Γ1 = 0.
As a numerical example, the approach is applied to simulate an embedded crack in a thin plate at tensile
loading. A transversely isotropic material with several fiber angles are simulated (see Fig. 2).
457
a) b) c)
d) e) f)
Figure 2. Comparison of crack propagation of an embedded crack in a thin plate at tensile load: a) initial crack
configuration, b) isotropic material and c-f transversely isotropic material (E1 /E2 = 3.0) with c) γ = 70◦ ,
c) γ = 75◦ , c) γ = 80◦ , c) γ = 90◦
3 Conclusion
The proposed phase-field approach based on Representative Crack Element (RCE) is applied to trans-
versely isotropic material. The derived fracture models can be equivalently used instead of discrete crack
models due to the general variational framework and consistent kinematics.
Acknowledgments
The authors gratefully acknowledge the financial support by the German Research Foundation (DFG)
within the priority program SPP 2020 (KA 1163/40).
References
[1] P. J. Blanco, P. J. Sánchez, E. A. Souza, and R. A. Feijóo. Variational Foundations and Generalized
Unified Theory of RVE-Based Multiscale Models. Archives of Computational Methods in Engineering
23 (2014), pp. 1–63. doi: 10.1007/s11831-014-9137-5.
[2] Alexander Schlueter. Phase Field Modeling of Dynamic Brittle Fracture. Technische Universität
Kaiserslautern, 2018.
[3] C. Steinke and M. Kaliske. A phase-field crack model based on directional stress decomposition.
Computational Mechanics 63 (2018), pp. 1019–1046. doi: 10.1007/s00466-018-1635-0.
[4] M. Strobl and T. Seelig. On constitutive assumptions in phase field approaches to brittle fracture.
Procedia Structural Integrity 2 (2016), pp. 3705–3712. doi: 10.1016/j.prostr.2016.06.460.
458

Reconstruction of a Ferroelectric Stripe Domain Array

due to Charge Defect Migration
∗1 2 1
Ivan Vorotiahin , Anna Morozovska and Yuri Genenko
1 Institute of Materials Science, Technische Universität Darmstadt, Otto-Berndt-Str.3, Darmstadt, Germany

2 Institute of Physics, National Academy of Sciences of Ukraine, pr. Nauky 46, Kyiv, Ukraine
Abstract.
Introduction
Field-driven migrating charged defects, such as impurities and oxygen vacancies, are often considered
as the widespread mechanism of aging in ferroelectric materials [2]. This process is usually studied
within a xed domain structure, where it is able to change charge distributions, however, it is known
that domain structure is actually aected by this process [1]. This work studies the evolution of a self-
organized polarization stripe-domain array induced by the movement of the oxygen vacancies throughout
an acceptor-doped ferroelectric lm. The processes are investigated by the Landau-Ginzburg-Devonshire
theory creating a self-consistent numerical model which thoroughly describes polar, electrostatic and
electromechanical properties of the lm including doping-induced semiconductor properties [3]. Due
to the eld-driven defect migration in the domain array, the array itself undergoes changes in its charge
balance which ultimately leads to domain shape modications. The processes exhibit several characteristic
times related to dierent mechanisms taking place inside the lm, among which are the defect migration
with the characteristic Maxwell time and the domain array response changing the domain conguration.
1 Problem Statement
A model of a 24-nm thin single-crystalline barium titanate lm was considered, xed on a metallic bottom
electrode with a compressive mist strain, with a free top surface exposed to the ambience (Figure 1). The
lm contains charged defects, namely, immobile acceptor impurities and mobile (donor) oxygen vacancies.
The domain structure of the lm arises spontaneously and rapidly, and then the migration of the mobile
defects becomes the main mechanism of the further self-organized polarization evolution. The goal of
this work is to establish the inuence of the defect migration on the ferroelectric properties of the lm.
The Landau-Ginzburg-Devonshire approach is chosen to describe polar, electrostatic, electromechanical,

and semiconductor properties of the lm, as well as the movement of charged defects througuout the
lm with time until its saturation. The Gibbs thermodynamic potential in relation to polarization,
electrostatic potential, mechanical stress, electron and hole, as well as donor and acceptor concentrations
comprises of a volume, a surface and dead-layer parts. The volume part consists of the Landau expansion,
gradient, electrostatic, elastic, semiconductor and defect terms:
GV = GLandau + Ggrad + Gel.stat + Gelast + Gsemi + Gdonor + Gacceptor (1)
A system of equations consists of the Landau-Khalatnikov equation, which is given by the variation of
the Gibbs free energy by polarization,
∗ Corresponding author: isvort@mm.tu-darmstadt.de
459
Figure 1. Schematic image of a homogeneously doped ferroelectric lm of thickness , with domains of width with
vertically oriented polarization (red and blue frames with arrows) and closure domains with laterally
oriented polarization (yellow arrows). The lm is clamped to a rigid conductive bottom electrode, has a
layer of surface screening charge on the top surface and is separated by a thick ambient layer of thickness
from the top electrode.
∂Pi δGV
Γ = (2)
∂t δPi
Poisson equation for electrostatic potential, with space charge created by electrons, holes, donors and
acceptors accounted for,
∂2ϕ ∂Pj
ε0 εb = + e(Zd Nd+ − Za − Na− + p − n) (3)
∂xi ∂xi ∂xj
a set of four Continuity equations for each charge specie with the general form of:
∂N 1 ∂JiN
= (4)
∂t Ze xi
and the generalized Hiike law with the elastic equilibrium condition to account for electromechanical
eects, such as electrostriction and exoeect
∂Pk
ui j = si jklσk l + Fi jkl + Qi jklPk Pl (5)
∂xl
∂σi j
=0 (6)
∂xi
all accompanied with boundary conditions ensuring blocking interfaces, grounded bottom electrode, mist
strain on the bottom surface and periodicity on the side biundaries -w and w.
2 Defect-Induced Processes
Figure 2 shows the time evolution of the vertical polarization, mobile donor concentration and charge
distribution over the lm at the acceptor doping level of 0.1 mol-percent. At the begining, donors
and acceptors distributed similarly and compensate each other's charge, however acceptors are immobile.
Donors, on the other hand, redistribute over time creating areas with excess and diminished concentrations
under the free top surface aligning themselves with the electrostatic potential. In response, polarization
domains start changing their shape creating upon saturation the wedged or trapezoidal forms with tilted
walls. Once walls are tilted, the space charge on them changes its sign.
460
Figure 2. Time evolution of the stripe domain structure in a 24 nm thick ferroelectric lm doped with a divalent
acceptor of concentration 0.1 mol-percent illustrated by snapshots of two-dimensional maps of vertical
polarization component (a-e), concentration of oxygen vacancies (f-j) and density of bound charge (k-
o). Numbers at the top and the bottom of plots (f-j) indicate the reached limiting values of donor
concentration.
Figure 3. Time evolution of the average surface potential in a 24-nm thick ferroelectric lm for a set of accep-
tor doping concentrations c0 from 0.01 mol-percent to 0.1 mol-percent. Inset: Zoom-in of the time
development from 10 to 106 seconds.
Figure 3 shows the time evolution of the average potential on the top surface at dierent acceptor
concentrations. The nonmonotonous shape of the curve indicates a competition between several processes
occuring in the lm. Comparing it to the distributions shown in Figure 2 it is possible to establish,
which processes are responsible for dierent behaviour of the curve parts. Charge redistribution process
dominates at characteristic times of 10
5 - 106 seconds and ultimately pulls the curve to the positive
potential. Appearing rst as the response, domain redistribution becomes dominant at larger timescales
and ultimately pulls the curve to the negative potential.
461
Acknowledgments
This work was supported by the Deutsche Forschungsgemeinschaft (DFG) via the grant No. 405631895
(GE-1171/8-1).
References
[1] Z. M. Fan and X. L. Tan. In-situ TEM study of the aging micromechanisms in a BaTiO3-based
lead-free piezoelectric ceramic. Journal of European Ceramic Society 38 (2018), pp. 34723477.
[2] Y. A. Genenko, J. Glaum, M. J. Homann, and K. Albe. Mechanisms of aging and fatigue in
ferroelectrics. Materials Science and Engineering: B 192 (2015), pp. 5282.
[3] A. N. Morozovska, E. A. Eliseev, Y. A. Genenko, I. S. Vorotiahin, M. V. Silibin, Y. Kim Y. Cao, M. D.

Glinchuk, and S. V. Kalinin. Flexocoupling impact on size eects of piezoresponce and conductance
in mixed-type ferroelectric semiconductors under applied pressure. Physical Review B 94 (2016),
p. 174101.
462

Phase-field simulation of the flexoelectricity in ferro-

electrics
Xian-Dong Zhou ∗1
and Bai-Xiang Xu1
1 Division Mechanics of Functional Materials, Department of Materials Science, TU Darmstadt
Flexoelectricity describes the linear coupling between the polarization and the strain gradient or the
coupling between the strain and the polarization gradient. Unlike other electromechanical coupling effects
such as piezoelectricity, which require the non-central symmetry of the structure, flexoelectricity applies to
all crystal symmetries. The flexoelectricity broadens the material selection arrange for electromechanical
application and has received recently significant research interest [3, 4].
The flexoelectricity in ferroelectrics can be characterized by nanoindentation tests, in which large strain
gradient is available. Most work focus on the estimation of the flexoelectric constants at room temperature
and few work discussed about nanoindentation test at high temperature. Recent nanoindentation tests
of BaTiO3 shows a super compressive elasticity of the material around Curie temperature, which means
a much larger indentation depth than that at room temperature under the same load. With increased
indentation depth, the distribution of strain as well as strain gradient changes around the indenter tip.
Therefore, it is interesting to understand how the flexoelectricity can affect the material behavior around
Curie temperature in nanoindentation tests.
In this work we extend a phase-field ferroelectric model to include the flexoelectricity [2, 5]. A weak form
formulation is developed. The formulation shows there are only first order derivatives in the weakest form,
therefore, conventional lower order h-type finite element can be adopted for numerical implementation.
The model is applied to study the nanoindentation tests of BaTiO3 ferroelectrics at high temperature,
particularly near Curie temperature. The reason why the material shows super compressive elasticity
property is analyzed by considering different possible boundary conditions and loading in nanoindenta-
tion tests. Different temperatures below and above the Curie temperature are also considered in the
simulation. In order to study how flexoelectricity affects the behavior of the material, a parametric study
is performed with the three indepedent flexocoupling constants for perovskite ferroelectrics with a cubic
parent phase. The domain evolution of the ferroelectrics during nanoindentation tests is analyzed with
and without flexoelectricity. A flexoelectric-induced paraelectric-ferroelectric phase transition slightly
above the Curie temperature is analyzed and compared with experimental results. Moreover, we sim-
ulated the flexoelectricity in nanoparticles with cation inhomogeneity and its induced strain gradient,
which is successfully applied to explain the experimentally observed star-like polar regions at extremely
high temperature [1].
References
[1] Leopoldo Molina-Luna, Shuai Wang, Yevheniy Pivak, Alexander Zintler, Héctor H Pérez-Garza,
Ronald G Spruit, Qiang Xu, Min Yi, Bai-Xiang Xu, and Matias Acosta. Enabling nanoscale flex-
oelectricity at extreme temperature by tuning cation diffusion. Nature communications 9.1 (2018),
p. 4445.
[2] D Schrade, R Mueller, BX Xu, and D Gross. Domain evolution in ferroelectric materials: A continuum
phase field model and finite element implementation. Computer methods in applied mechanics and
engineering 196.41-44 (2007), pp. 4365–4374.
[3] AK Tagantsev. Piezoelectricity and flexoelectricity in crystalline dielectrics. Physical Review B 34.8
(1986), p. 5883.
∗ Corresponding author: x.zhou@mfm.tu-darmstadt.de
463
[4] Bo Wang, Yijia Gu, Shujun Zhang, and Long-Qing Chen. Flexoelectricity in Solids: Progress, Chal-
lenges, and Perspectives. Progress in Materials Science (2019).
[5] Shuai Wang, Min Yi, and Bai-Xiang Xu. A phase-field model of relaxor ferroelectrics based on
random field theory. International Journal of Solids and Structures 83 (2016), pp. 142–153.
464
MS 23: Non–Newtonian Fluids

Comparison of different rheological models in pressure-

hole error simulation
1
Hogenrich Damanik ∗
and Stefan Turek1
1 Lsiii Angewandte Mathematik, TU-Dortmund
Abstract. We present numerical techniques for the calculation of steady viscoelastic fluids given in differential
form and perform numerical simulations of the hole-pressure error of different rheological models. The numer-
ical techniques are based on a well-developed mixed Finite element approach for steady fluid flows which can
deal with complex rheological models in a black-box manner. We consider rheological models from quasi Newto-
nian fluids, i.e with nonlinear viscosity function, as well as models from quasi linear viscoelasticity. The numer-
ical simulation shows typical non-symmetry streamlines and a secondary vortex when increasing the viscoelas-
ticity.
Introduction
Hole-pressure error is an interesting test problem for both numerical as well as experimental works because
it has a simple geometry which can be easily fabricated and attached at the end of an extrusion process.
Furthermore, there exists a theoretical work for an analytical expression of hole-pressure error, see [6],
too. Hole-pressure error deals with the wrong pressure reading from the instrument due to drilling of the
fluid chamber/channel. Instead of giving the actual pressure at a given point of location on the wall of
the channel, a wrong pressure is measured because the tip of the pressure stick/transducer does not close
the channel wall appropriately. The same situation might happen to several other points of measurement,
and thus giving an inaccurate data reading. In this regard, numerical simulation can be used to provide
first normal stress difference data up to a small range of elasticity.
1 Rheological models
We consider several rheological models for the numerical simulation. An incompressible steady flow of
generalised Newtonian fluid needs to satisfy
(
(u · ∇)u = −∇p + ∇ · T
(1)
∇·u=0
q
with the constitutive law T = 2ηs (γ̇, p)D, shear rate γ̇ = tr(D2 ), and symmetric velocity gradient
D = 21 (∇u + ∇uT ). In the numerical context, the above Eq. (1) is well-known as velocity-pressure
formulation (u, p). The fluid viscosity function can be written as, i.e., the power law model
r
ηs (γ̇) = η0 ( + γ̇ 2 )( 2 −1) (2)
with r is the constant parameter index that describes whether the fluid is in the region of shear thinning
( 2r < 1) or shear thickening ( 2r > 1). While there exists no normal stress difference for Newtonian fluids,
we consider also fluids with elastic properties, such there is an elastic contribution to the constitutive law
T = 2ηs (γ̇, p)D + σ. (3)

This additional elastic properties must be calculated/integrated along the historical path of each fluid
particle, leading to the so-called integral model of viscoelastic fluids. Since the path for integration is
∗ Corresponding author: hdamanik@mathematik.tu-dortmund.de
465
unknown in the Eulerian frame work, Oldroyd in [8] and Giesekus in [4] provide differential forms of
σ-equation for quasi linear models where time and strain can still be separated. For numerical purposes,
these original models are often reformulated. We consider the logarithm of the conformation stress τ
reformulation (LCR)
ηp
τ = (σ − I), ψ = log(τ ) (4)
Λ
with fluid viscosity ηp and fluid relaxation time Λ. The Oldroyd-B and Giesekus model can now be
written as in the previously numerical work of [3, 7]:
∂ψ 1
+ (u · ∇)ψ − (Ωψ − ψΩ) − 2B = f (ψ), f (ψ) = (e−ψ − I) − αeψ (e−ψ − I)2 (5)
∂t Λ
where parameter α = 0 gives Oldroyd-B model, and α 6= 0 gives Giesekus model. The two models come
from the kinetic theory. Other models which are also from the same theory are FENE-P and FENE-CR.
In the FENE-P and FENE-CR model the maximum polymer extension is bounded by some value L, see
[1]. It can be written in LCR form as
∂ψ 1
+ (u · ∇)ψ − (Ωψ − ψΩ) − 2B = − f (R, L)I − αf f (R, L)e−ψ (6)
∂t Λ
where f (R, L) = 1
2 , αf = 0 for FENE-P and αf = 1 for FENE-CR, and where R2 = tr(τ ).
1− R
L2
2 Numerical treatment
We solve the Eq. (1) for generalised Newtonian fluids, and solve the Eq. (1) and (5) or (6) for viscoelastic
fluids. In this regard, finite element method is employed for space approximation. Consequently, we
have to choose approriate finite element pairs which is subject to the well-known compatibility condition
between the velocity and pressure space for Newtonian fluids, see [5], and between velocity-pressure and
stress for viscoelastic case, see [2]. We choose the finite element pair Q2 P1 with the following definition:

Q2 (T ) = q ◦ ψT−1 : q ∈ span < 1, x, y, xy, x2 , y 2 , x2 y, y 2 x, x2 y 2 > (7)
with nine local degrees of freedom located at the vertices, midpoints of the edges and in the center of the
quadrilateral. The space P1 (T ) consists of linear functions defined by

P1 (T ) = q ◦ ψT−1 : q ∈ span < 1, x, y > (8)
with the function value and both partial derivatives, located in the center of the quadrilateral, as its three
local degrees of freedom. Then, the interpolation functions are
Vh = {ṽh := (uh , τ h ) ∈ X(Ωh ), ṽh|T ∈ [Q2 (T )]6 ∀T ∈ Th }, (9)
Qh = {ph ∈ L20 (Ωh ), ph|T ∈ P1 (T ) ∀T ∈ Th }. (10)
Finally, the discrete finite element formulation can be written as a nonlinear saddle point problems,
namely
A BT ũ rhs ũ
= (11)
B 0 p rhs p
| {z }
Nonlinear operator A
where ũ consists of the coefficient vector representing the velocity and LCR tensor. Here, A includes
all convection and diffusion operators. We employ Newton iteration method to the nonlinear system of
Eq. (11) and use finite differencing Jacobian as a black-box for the linearisation of Newton, see [3] for
implementation aspect.
3 Numerical discussion
It is not our intention to simulate a particular fluid, so the following quantities (Tab. 1) would be used
as a prototypical test problem. The degree of elasticity is controlled by Weissenberg number We = Λγ̇w
where γ̇w is calculated on the channel wall.
A typical streamline of generalised Newtonian fluid for Hole-pressure error simulation is shown in Fig. 1,
where one observes symmetric patterns as a typical lack of elasticity in the fluid model, although power
law model with r = 1.5 lies in the region of shear thinning fluid.
466
Table 1. Material parameters

Model Power law Old-B Gie Fene-p Fene-cr
Para- r = 1.5 α=0 α = 0.01 α = 0, α = 1.0,
meter L2 = 100 L2 = 100
Figure 1. Streamlines
Acknowledgments
The AIF/IGF-project 18485N is gratefully acknowledged.
References
[1] Bird, R. B., Curtiss, C. F., Armstrong, R. C., and Hassager, O. Dynamics of polymeric fluids. Vol.
2, New York. Wiley, 1987.
[2] Brezzi, F. and Fortin, M. Mixed and Hybrid Finite Element methods. Springer, Berlin, 1986.
[3] Damanik, H., Hron, J., Ouazzi, A., and Turek, S. A Monolithic FEM approach for the log-conformation
reformulation (LCR) of viscoelastic flow problems. Journal of non-Newtonian Fluid Mechanics 165
(2010), pp. 1105–1113.
[4] Giesekus, H. A simple constitutive equation for polymer fluids based on the concept of deformation-
dependent tensorial mobility. Journal of non-Newtonian Fluid Mechanics 11 (1982), pp. 69–109.
[5] Girault, V. and Raviart, P. A. Finite Element Methods for Navier-Stokes equations. Berlin-Heidelberg.
Springer, 1986.
[6] K. O. Higashitani and W. G. Pritchard. A kinematic calculation of intrinsic errors in pressure
measurements made with holes. Trans. Soc. Rheo. 16 (1972), p. 687.
[7] Hulsen, M. A., Fattal, R., and Kupferman, R. Flow of viscoelastic fluids past a cylinder at high
Weissenberg number: Stabilized simulations using matrix logarithms. Journal of Non-Newtonian
Fluid Mechanics 127 (2005), pp. 27–39.
[8] Oldroyd, J. G. On the formulation of rheological equations of state. Proc. R. Soc. London Ser. A
200 (1950), pp. 523–541.
467
468

Shape Optimization in Polymer Extrusion Using Shear-

Thinning Fluids
1
Sebastian Eusterholz ∗
and Stefanie Elgeti1
Abstract. Extrusion is a manufacturing process suitable for endless profiles with a fixed cross-section and is
widely used in plastics processing. One key component of this process is the extruder: Its purpose is to melt,
transport, and homogenize the raw material. To produce a continuous melt flow, solid polymer particles are
taken in and then molten due to shear-heating as they are conveyed through the extruder. Despite other kinds
of extruders also being utilized, this work focuses on single-screw extruders (SSEs). While cheap in design and
operation, SSEs per se have a – compared to, e.g., twin-screw extruders – reduced mixing ability. Therefore, the
resulting products often suffer from a lack of homogeneity in both temperature and material distribution. This
is especially relevant if additives are introduced into the melt. As a means to increase homogeneity, it is com-
mon in the plastics processing industry to introduce flow obstacles into the screw’s end section. This enhances
the reordering of the essentially laminar flow. The problem with this approach is that while the mechanisms of
laminar mixing are mostly understood, there is no rigorous design principle relating the mixing element geom-
etry to certain flow characteristics. Furthermore, since single-screw extruders have to be operated at high tem-
peratures and pressures, and thus, transparent housings for experimental studies of the flow inside those ma-
chines cannot be used; experimental data is very limited. It is our aim to replace the current, experience-based
design process with shape-optimization methods. The proposed concept comprises three steps: (1) Computa-
tion of the flow field, (2) computation of the associated objective function value, and (3) determination of op-
timized shape parameters. While the last point is dealt with by the utilized optimization framework, this work
discusses two aspects: (1) We detail on geometry modification techniques to alter given mixing-element shapes
retaining boundary-conforming meshes, and (2) we review how such methods in combination with fast simula-
tions of shear-thinning fluids can be used for the shape-optimization of industrially relevant mixing elements.
Introduction
Numerical shape optimization as a design aid has had an impact on the plastics processing industry
already. The work of Elgeti et. al [1] has proven its potential to support design in areas where experimental
data is rare and shape is crucial to the product’s quality. A similar situation is observed in the field of
mixing in single-screw extruders (SSEs).
SSEs are frequently utilized in extrusion due to their low operating and investment costs. By design,
single-screw extruders, unfortunately, provide only limited mixing ability which is a likely cause for
insufficiently homogeneous products. Plastics manufacturer have addressed that problem by adding flow
obstacles – the mixing elements – to the rotating downstream tip of the screw to enhance reorientation
within the flow. An exemplary configuration is shown in Fig. 1. Interestingly, there is no rigorous design
Figure 1. Exemplary SSE screw. The screw’s tip is shown in purple with mixing elements of typical rhomboid
shape highlighted in yellow.
pattern – also because of a lack of experimental data on the flow inside extruders. This work attempts to
∗ Corresponding author: eusterholz@cats.rwth-aachen.de ORCID iD: ORCID 0000-0001-7795-9760
469
take a first step to bridge that gap by numerical design using shape optimization. Following the approach
presented in [1], we propose methods to extend the concepts utilized in die design to the numerical design
of mixing elements.
The numerical design of mixing elements by shape optimization has two main building blocks: (1)
A forward simulation of the physical phenomena describing mixing in extruders, and (2) the actual
optimization framework. In the following sections we will first address the latter and start by proposing
a technique to manipulate the computational mesh to reflect changes in the geometry. Following up on
this, our setup for the flow computation of shear-thinning polymer melts is presented.
1 Geometry manipulation using free-form deformation and elasticity
The approach proposes that the update of a given geometry, and subsequently the manipulation of the
mesh is composed of three steps: (1) Modification of a given spline representation of the geometry, (2)
according manipulation of the boundary mesh, and (3) update of the volume mesh. Starting point of
the proposed method is a spline description of some initial mixing-element geometry – as given from
Computer-Aided Design. A computational domain Ω with boundary Γ is constructed around that geom-
etry, representing a portion of the extruder’s flow channel. One subdomain of Γ, called Γspline , coincides
with the input geometry.
Manipulating the boundary mesh – the final goal to reflect shape changes in the simulation – is done by
deforming the input spline first and then moving the mesh nodes accordingly. By accordingly we mean
that mesh nodes lying on the input spline are moved such that they remain on the spline with no change
in their parametric coordinate.
Previous works utilizing a similar approach rely on modifying the boundary spline by directly shifting
its control points [1]. Since this approach poses restrictions on the usable input splines, we extend the
method by manipulating the spline using free-form deformation. The concept, as introduced by Sederberg
and Parry [3], is outlined in Fig.2.
SCS (ξ, η) SĈS (ξ, η)
VD◦M(CV ) (ξ, η, ζ)
VCV (ξ, η, ζ) SVD◦M(C (V −1 (CS )) (ξ, η)
V)
s.t. CS ⊆ CV
Figure 2. Free-from deformation. A volume spline VCV (ξ, η, ζ) in R3 is constructed around a given surface spline
SCS (ξ, η), such that V ’s control polygon CV encloses the surface spline’s control polygon CS . Deformation
to the enclosed body is imposed by applying a deformation D to a subset of V ’s control points, masked
via M. The deformed configuration of SCS (ξ, η) is obtained by evaluating the position of the points of
CS in the deformed configuration of V .
Once the boundary spline has been modified, and the boundary mesh nodes are shifted to the new
positions, it is advantageous to update the volume mesh to retain mesh quality. We use the elastic-mesh-
update method described in [2]. The deformed configuration is computed, treating the volume mesh as
an elastic body that reacts to displacements imposed on its boundary. The governing equations, with the
470
displacement z being the unknown, are given in Eqs. (1) to (3).
∇ · σ mesh = 0 , (1)
σ mesh (z) = λmesh (tr εmesh (z)) I + 2µmesh εmesh (z) , (2)
1
εmesh (z) = ∇z + (∇z)T . (3)
2
We partition the boundary into three parts as Γ = Γspline ∪Γf ixed ∪Γslip . Tangential movement is allowed
for on Γslip , whereas we restrict any movement on Γf ixed , and prescribe the displacement obtained from
the spline deformation on Γspline . With this, we avoid any remeshing but still obtain a boundary-
conforming mesh.
2 Simulation of shear-thinning flows
Given the deformed mesh we obtain the flow and temperature fields by solving steady incompressible
Navier-Stokes equations strongly coupled to the temperature equation. We apply strong coupling between
flow and temperature problem due to the material law, for which we choose the Carreau model in the
following form:
A
η= C
, (4)
(1 + B γ̇)
and apply a WLF temperature correction as first proposed by Williams et al. [5] and generalized by van
Krevelen [4] as:
8.86·(TM −TRef ) 8.86·(T −TRef )
−
η0WLF (T ) = η0 · e 101.6+TM −TRef 101.6+T −TRef . (5)
In Eq. (4), A is equal to the viscosity at zero shear η0 , B determines the transition to the shear-thinning
regime and C describes the double logarithmic decay of η with increasing shear rates. In Eq. (5), TM
and TRef denote the measurement and reference temperatures. Following [4], TRef is chosen to be 43 K
above the glass transition temperature. Pressure corrections are neglected so far.
3 Shape optimization in view of enhanced distributive mixing
We attempt to optimize the mixing element’s geometry with regard to improved distributive mixing
capabilities. Thus, we aim to maximize the re-orientations imposed by the mixing element. Instead
of relying on particle tracking methods, we solve a set of n scalar concentration problems ci , where
1 ≤ i ≤ n, to estimate the material distribution within the flow domain. Given Γin and Γout as the inflow
and outflow boundaries of the computational domain, we then compute the objective function J as:
J = k∆dk2 with ∆di = kP (∆xi )k2 = kP (xi,max, in − xi,max, out )k2 (6)
where xi,max, in/out denotes

the position of maximum concentration at Γin/out , i.e. xi,max, in/out ∈ Γin/out ,
such that ci xi,max, in/out ≥ ci (x) ∀x ∈ Γin/out . To obtain a proper scaling, the mapping P ensures that
only the displacement projected onto Γin/out is taken into account, such that P (∆xi ) · n = 0 holds with
n being the normal vector on Γin/out . It should be noted that this approach is directly applicable only
to cuboid domains.
4 Summary
A new approach has been presented that enables shape optimization for mixing elements. Instead of
defining one monolithic problem, a partitioned approach was chosen for both, mesh adaptation and flow
computation. With the presented approach we are able to design new enhanced mixing elements by
applying shape optimization to mixing elements given an arbitrary initial input geometry.
471
Acknowledgments
The authors gratefully acknowledge the support of DFG provided by grant “Automated design and
optimization of dynamic mixing and shear elements for single-screw extruders” (EL 745/5-1)
References
[1] S. Elgeti, M. Probst, C. Windeck, M. Behr, W. Michaeli, and Ch. Hopmann. Numerical Shape
Optimization As an Approach to Extrusion Die Design. Finite Elem. Anal. Des. 61 (2012), pp. 35–
43. doi: 10.1016/j.finel.2012.06.008.
[2] A.A. Johnson and T.E. Tezduyar. Mesh Update Strategies in Parallel Finite Element Computations
of Flow Problems with Moving Boundaries and Interfaces. CMAME 119 (1994), pp. 73 –94.
[3] T. W. Sederberg and S. R. Parry. Free-form Deformation of Solid Geometric Models. SIGGRAPH
Comput. Graph. 20.4 (1986), pp. 151–160. doi: 10.1145/15886.15903.
[4] D. W. Van Krevelen and K. Te Nijenhuis. Properties of polymers: their correlation with chemical
structure; their numerical estimation and prediction from additive group contributions. Elsevier, 2009.
isbn: 978-0-08-054819-7.
[5] M. L. Williams, R. F. Landel, and J. D. Ferry. The Temperature Dependence of Relaxation Mecha-
nisms in Amorphous Polymers and Other Glass-forming Liquids. Journal of the American Chemical
Society 77.14 (1955), pp. 3701–3707. doi: 10.1021/ja01619a008.
472

Heat Dependent Thermo-Viscous Shear-Thinning Fluid

Flowing around a Double-Sphere Configuration within
a Pipe
∗ 1,2
Markus Rütten
1 DLR - German Aerospace Center, Institute of Aerodynamics and Flow Technology, Bunsenstrasse 10, 37073 Göttingen, Germany
2 University of Kassel, Dept. of Mechanical Engineering, Institute of Fluid Mechanics, 34125 Kassel, Germany
Abstract. Non-Newtonian fluid properties can change the flow behaviour and the resulting flow topology
significantly, even for low Reynolds number laminar flows. The prediction of the flow field is even more
difficult when thermal effects come into play. In particular when high shear rates occur and temperature
is changing the fluid properties, unexpected flow phenomena can be observed. Focal point of this study
is the laminar pipe flow of a thermo-viscous shear thinning Carreau fluid around double-sphere configu-
rations. The two spheres are arranged in-line in pipe axial direction. In dependence of the distance be-
tween the spheres the flow topology around the sphere pair changes. The non-isothermal non-newtonian
shear-thinnning fluid flows reveal temperature dependent deflections of the wake flow. Complicated flow
separation and attachment structures between and behind the spheres occur. In the present CFD study
sphere-wall temperature-dependent flow-structures are investigated. The temperature dependence of the fluid
viscosity is modelled by applying the standard Williams-Landel-Ferry equation. It is shown that thermal-
viscous shear thinning due to high shear rates and heat transfer from the sphere walls leads to a signif-
icant decrease of viscosity of three orders of magnitude in the near-field of the spheres. The locally re-
duced viscosity can be used to control the flow around the double sphere flow configuration. In the present
simulation campaign thermal boundary conditions are varied over a wide range in order to derive rela-
tions between heat transfer, non-newtonian fluid properties and the behaviour of the double-sphere flow.
Introduction
Non-Newtonian fluids reveal a wide variety of properties due to their non-linear behaviour of viscosity in
shearing flows. Since these types of fluids are widely used in technical, industrial and medical applications
and also occur in biological flows, the knowledge of their complex properties and flow behaviour is
necessary for the design of novel devices. The complex flow behaviour is caused by their molecular
composition. Long entangled molecular chains lead to higher viscosities. High shear flow rates can
tighten and align the molecular chains which leads to a decrease of the shear viscosity. This effect is
called shear-thinning. The resulting flow phenomena of these processes are not easily to predict, in
particular in complex flow applications [2], [3]. The flow prediction is even more difficult when additional
thermal loads are playing a significant role, since middle and high-molecular-dense fluids are sensitive to
the temperature influences on the viscosity [9]. Thus, heat transfer at walls, its effect on the non-linear
viscosity and the resulting flow phenomena are a main focal point in the research field of non-Newtonian
fluids [8], [5]. In particular, heat exchanger, lubrication systems or oil coolers are technical applications
in which the properties of non-Newtonian fluids are essential for the design and layout. Since numerical
simulations, conducted during the design process for such applications, have to be based on very accurately
described non-Newtonian fluid models [7], experimental and numerical studies are necessary to capture
all phenomena, which are related to thermal and shear-thinning viscosity properties of these fluids, in
order to improve the applied fluid models. In this context the non-Newtonian flow of a shear-thinning
fluid around a double-sphere configuration within a pipe is investigated.
∗ Corresponding author: markus.ruetten@dlr.de
473
1 The Non-Newtonian Fluid Model and the Double-Sphere Flow Case
In contrast to Newtonian fluids, the properties of non-Newtonian fluids play a crucial role in shearing
flows due to their ability to change their viscosity by orders of magnitude in dependence of local shear
rates. If thermal loads also play a role, then heat transfer can enforce the viscosity change of those fluids.
Thus, the accurate modelling of the non-Newtonian fluid properties of the considered thermo-viscous
shear-thinning fluid is crucial for the description of those flow problems, which is demonstrated in the
present simulation campaign. Fluid viscosity modelling together with a suitable CFD grid and a strict
numerical simulation strategy is a precondition for reliable simulation results.
In this study thermo-viscous shear-thinning fluids without any yield stress q or temporal relaxation prop-
erties are considered: Both the temperature T and the shear rate, γ̇ = 12 [D : D], have an impact on
the viscosity of these class of non-Newtonian fluids. The simplest shear-thinning fluid is the Ostwald - de
Waele fluid [6], whose viscosity is modelled by a power law. However, fluid models purely based on the
power law have the model limitation that they do not possess any lower and upper viscosity restrictions.
This leads to fluid modelling errors at very low and very high shear rates, where the shear dependent
viscosity reaches unrealistic values. Moreover, numerical simulations reveal that Power-Law fluids show
worse convergence rates, in particular
in flow regions, in which the second invariant I2 of the strain
tensor D = 21 grad~v + grad~v T is small, as, for example, in nearly uniform flows. An alternative to a
Power-Law fluid is the Carreau fluid model [4], which has the advantage to have both a constant zero
viscosity and a constant infinite viscosity. With this model the properties of real fluids are reflected very
well [2]. The flow curve of a Carreau fluid is calculated by using:
n−1
η = η∞ + (η0 − η∞ ) 1 + (λγ̇)2 2 . (1)
In equation (1) n is the flow index, η0 the lower limit or zero viscosity, η∞ the upper limit or infinite
viscosity and γ̇ the generalised shear rate. The parameter λ controls the transition from the Newtonian
to the non-Newtonian fluid behaviour. This parameter can be determined by the characteristic stress
τγ , at which the deformation-viscosity behaviour of the fluid changes: λ = η0 /τγ . In principle, this
time constant is the reciprocal value of the critical shear rate γ̇c , at which exceeding shear rates start to
reduce the viscosity of the fluid. In this simulation study we used the following fluid parameters, n = 0.2,
λ = 0.5 s, η0 = 10 Pas, η∞ = 0.0002 Pas, which are describing the fluid properties valid for a reference
temperature Tref = 403.15 K. The associated flow curve is depicted in Fig. 1. The small values of the
time constant and the flow index are consciously chosen in order to achieve a high sensitivity of the fluid
viscosity in particular to lower shear rates, so that the viscosity can drop orders of magnitude over the full
range of the expected flow shear rates. This is not non-physical, since low-density polymers, carbon and
silicone oils or especially synthesised engine oils just reveal such fluid properties. The reference density
of the fluid is ρref = 1000 kg/m3 . The density changes only slightly with temperature.
101 Carreau fluid at reference temperature

heated Carreau fluid - aT = 0.6
heated Carreau fluid - aT = 0.4
100
η(γ̇) [P a · s]
10−1
10−2
10−3
10−4
10−1 101 103 105 107 109 1011

γ̇ 1s
Figure 1. Fluid curves for a Carreau fluid model, n = 0.2, λ = 0.5 s, η0 = 10 Pas, η∞ = 0.0002 Pas, at reference
temperature and for temperatures leading to viscosity-temperature shift factors of 0.6 and 0.4.
474
Modelling thermal-viscosity is a research field on its own and many temperature dependent viscosity
models exist. However, there is an observation that for a certain group of non-Newtonian fluids, low and
high density polymers or certain silicone oils, a change of the global temperature level only leads to a
shift of the flow curve, whereas the basic shape of the flow curve is barely changed. This observation has
led to a certain modelling concept: a so called master flow curve is measured for a reference temperature,
thereafter, this master flow curve is adapted to the new temperature level by a viscosity-temperature shift
factor aT . This shift factor is called time-temperature superposition parameter, following the notation
of [9], it is also called Andrade factor [1]. Now modelling the temperature influence on the viscosity is
reduced to determine this factor. Different equations can be applied to this problem. In this work the
so-called William-Landel-Ferry equation (WLF-equation) is used to determine the viscosity-temperature
shift factor [9], [5]. The WLF-equation reads as follows:

η(γ̇, T ) −C1 (T − Tref )
log (aT ) = log = . (2)
η(γ̇, Tref ) C2 + (T − Tref )
In this equation, Tref is the reference temperature, for which, in this study, a value of 403.15 K is
assumed. The fluid parameters C1 and C2 have to be determined experimentally for each non-Newtonian
fluid, however, in this campaign the common values for polymers, C1 = 8.86 and C2 = 101.6 K, are used.
Eventually, the thermal-viscosity model is combined with the Carreau fluid model equation leading to:
aT (η0 − η∞ )
η (γ̇, T ) = aT η∞ + 1−n . (3)
2 2
1 + (aT λγ̇)
The flow around two spheres within a pipe is considered. The spheres are placed on the middle axis of
the pipe and they are arranged in-line, streamdown in pipe axial direction. The pipe has a lenght of
1000 mm an a diameter D of 200 mm. The first sphere is located 300 mm streamdown of the pipe entry.
The diameter d of the spheres is 60 mm. The distance between the spheres has been varied from 2d to 4d
in 0.5d steps. In this work only the even cases are discussed. CFD grids have been generated using the
commercial mesher “Centaur” from CentaurSoft. Besides the pipe inflow and outflow all boundaries are
treated as viscous walls. The number of grid points and cells depends on the sphere distance, however,
on average the hybrid grid consists of 18 Mio points, 15 Mio prisms, 58 Mio tetraeders, 0.5 Mio triangles
and 20000 quadrilateral grid cells. The prisms are forming a structured grid close to the wall, which
consists of 32 prism layers. This high number of layers is necessary for resolving both, the momentum
driven boundary layer flow and also the thermal boundary layer. In order to capture the velocity profile
slopes accurately the first wall normal cell height is on average 0.0032 mm, the associated growth rate in
wall normal direction is 1.12.
The numerical simulations have been performed using the DLR THETA code for solving the incom-
pressible Navier-Stokes equations. This finite volume code utilizes the so-called collocated respectively
non-staggered dual grid approach storing all flow variables at the same location. An efficient coupling
of the velocity and the pressure fields is ensured by using the Chorin projection method or the SIMPLE
algorithm. For spatial discretisation different variations of 2nd and 3rd order upwind methods as well as
2nd order central schemes can be used. Here, a 3nd order scheme was chosen in order to get very smooth
data for calculation of derived vector fields. The well-known checkerboard instability of the pressure is
eliminated by a 4th order stabilisation term, which is added to the left and the right hand side of the
Poisson equation. A matrix-free formulation is used for solving the linear equations. This formulation
reduces the memory requirements of the code considerably. A variety of multi-grid methods ensures
efficient solutions of the linear equations, even on fine grids. Domain decomposition is used as paralleli-
sation concept. The parallel efficiency is high on desktop computers with only a few CPUs as well as on
massive parallel systems with thousands of CPUs. In the considered flow case a 4-w multi-grid cycle has
been performed to accelerate the CFD calculations. The projection method was used for the unsteady
calculation, in which the time-step was set to 0.001 s resulting in grid cell CFL numbers less than one.
The three point backward scheme, a variation of a BDF2 scheme, was used for temporal discretisation.
As inflow condition a velocity block profile with 3.5 m/s was set. At the entry the local Reynold number
is 70, based on the pipe diameter, zero viscosity and the reference density. The entering fluid has the
reference temperature of 403.15 K. During the simulation campaign the wall temperature of the first
sphere was increased from 403.15 K to 453.15 K in 10 K steps. In order to capture the impact of the heat
transfer on the fluid viscosity equation (3) was implemented in the DLR-THETA code.
475
2 Results
The following images show the flow simulation results for three different sphere-distance cases, as men-
tioned above. For each only the results for the isothermal flow case with the reference temperature as
the wall temperature of the first sphere and the results with the highest wall temperature of 453.15K are
shown. For all cases contour plots of the local shear rate on the spheres and of the velocity magnitude
on the mid-sectional pipe cut-plane are shown. Additionally streamlet visualisations of the flow on the
cut-plane are also depicted in order to reveal local flow topologies. The lowest sphere-distance case reveals
Figure 2. Visualisation of flow structures for case 1, wall Figure 3. Visualisation of flow structures for case 1, wall
temperature of the first sphere 403.15 K temperature of the first sphere 453.15 K
a strong coupling of double-sphere flow: in the isothermal flow case the second sphere is fully attached
to the wake flow of the first sphere. Its symmetrical wake vortex hits the second sphere and generates
an opposite shear flow at the front part, see Fig. 2. Interestingly the wake flow separation of the second
sphere is barely affected by the onflow. The increase of the wall temperature of the first sphere leads to
delayed separation and, hence, to a confined asymmetric wake flow between the spheres. As depicted in
Fig. 3, this results in spiraling shear stress structures on the back side of the first and the front side of the
second sphere. Even the wake flow of the second sphere shows asymmetries. With increasing distance of
the spheres the flow field interaction reduces. The middle case reveals an induced asymmetry of the first
wake flow by the second sphere, although the flow in the near field of the first sphere is quite symmetric.
The shear stress pattern of the second sphere reveals an unsteady spatially oscillating structure which is
the foot print of the onflowing wake field of the first sphere, shown in Fig. 4. Despite this strong temporal
and spatial characteristic the wake flow of the second sphere is nearly symmetric. In case of the heated
wall the confined wake flow does only barely affect the flow around the second sphere, see Fig. 5. In the
last sphere-distance case for both wall temperatures the flow fields around the spheres are more or less
decoupled, compare the Fig. 6 and 7. The increased wall temperature of the first sphere leads again to
the unstable confined wake observed before. Only a slight asymmetry is visible on the front side of the
second sphere.
476
3 Conclusion
In this work the flow around two spheres in a pipe was investigated. It could be observed that, when
high shear rates occur and additionally heat is transferred, the fluid properties and, thereupon, the flow
topology between and behind the spheres is changing drastically. The thermo-viscous shear-thinning effect
of the Carreau fluid leads to a narrowed wake flow behind the first sphere, but also to a significant increase
of flow instabilities and, eventually, vortical flow asymmetries. These effects are strongly dependend on
the distance between the spheres. The complicated flow separation and attachment structures between
and behind the spheres are a focal point of further research work.
Acknowledgments
The author would like to thank the German Aerospace Center (DLR) and the University of Kassel for
greatly supporting this research work.
References
[1] E.N.D. Andrade. A theory of the viscosity of liquids. Part 1. Philosophical Magazine 17 (1934),
pp. 497–511.
[2] R.B. Bird. Non-Newtonian behavior of polymeric liquids. Physica A: Statistical and Theoretical
Physics 118.1–3 (1983), pp. 3–16.
[3] R.B. Bird and O. Hassager. Dynamics of Polymeric Liquids: Vol.1: Fluid Mechanics. Dynamics of
Polymeric Liquids. London, New York: John Wiley & Sons, 1987.
[4] P.J. Carreau. Rheological equations from molecular network theories. J. Rheol. 16.1 (1972), pp. 99–
127.
[5] J.D. Ferry. Viscoleastic Properties of Polymers. Third. New York: John Wiley & Sons, 1980.
[6] W. Ostwald. Über die rechnerische Darstellung des Strukturgebietes der Viskosität. Kolloid-Z 47.2
(1929), pp. 176–187.
[7] R.G. Owens and T.N. Phillips. Computational Rheology. Computational Rheology. Imperial College
Press, 2002. isbn: 9781860941863.
[8] A.H.P. Skelland. Non-Newtonian flow and heat transfer. New York: John Wiley & Sons, 1967.
[9] M. L. Williams, R. F. Landel, and J. D. Ferry. The Temperature Dependence of Relaxation Mecha-
nisms in Amorphous Polymers and Other Glass-forming Liquids. Journal of the American Chemical
Society 77 (1955), pp. 3701–3707.
477
478
Authors Index
A E
Abali, Bilen Emek . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Ebna Hai, Bhuiyan Shameem Mahmood . . . 319

Abdusalamov, Rasul . . . . . . . . . . . . . . . . . . . . . . . 219 Eckert, Christoph . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Abedi, Masoud . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 Elgeti, Stefanie . . . . . . . . . . . . . . . . . . . . . . . . 331, 469
Adámek, Vítezslav . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Ellmer, Franz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Aldakheel, Fadi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 Eusterholz, Sebastian . . . . . . . . . . . . . . . . . . . . . . 469
Alkassar, Muhannad . . . . . . . . . . . . . . . . . . . . . . . 389
Arbeiter, Florian . . . . . . . . . . . . . . . . . . . . . . . . . . . 371 F
Arnold, Martin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Aygün, Serhat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Fan, Meng . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Fangye, Yannick Francis . . . . . . . . . . . . . . . . . . . 223
B Felger, Julian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
Fiedler, Melanie . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Balzani, Daniel . . . . . . . . . . . . . . . . . . . . 27, 223, 405 Fiedler, Robert . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Bartel, Thorsten . . . . . . . . . . . . . . . . . . . . . . . . . . . 327 Fleck, Michael . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
Barthold, Franz–Joseph . . . . . . . 55, 283, 343, 375 Flessing, Lars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Bause, Markus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319 Fohrmeister, Volker . . . . . . . . . . . . . . . . . . . . . . . . 323
Becker, Marvin . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377 Frey, Christian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Becker, Wilfried . . . . . . . . . . . . . . . . . . . . . . . 359, 363 Frischkorn, Jan . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
Beer, Michael . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 Fritzen, Felix . . . . . . . . . . . . . . . . . . . . . . . . . . 417, 421
Behr, Marek . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
G
Berezovski, Arkadi . . . . . . . . . . . . . . . . . . . . . . 15, 23
Berthold, Christian . . . . . . . . . . . . . . . . . . . . . . . . 311 Göbel, Luise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
Bianco, Marcelo J. . . . . . . . . . . . . . . . . . . . . . . . . . 291 Gabriel, Dusan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Biehl, Johanna K. . . . . . . . . . . . . . . . . . . . . . . . . . 315 Ganzosch, Gregor . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Bonari, Jacopo . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 Gebhardt, Clemens . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Brands, Dominik . . . . . . . . . . . . . . . . . . . . . . . . . . . 235 Genenko, Yuri. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .459
Bremicker-Trübelhorn, Sascha . . . . . . . . . . . . . 295 Gierden, Christian . . . . . . . . . . . . . . . . . . . . . . . . . 413
Brummund, Jörg . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 Gilbert, Robert P.. . . . . . . . . . . . . . . . . . . . . . . . . .401
Budday, Dominik . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Gilbert, Rose Rogin . . . . . . . . . . . . . . . . . . . . . . . 191
Budday, Silvia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Glatzel, Uwe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
Bulling, Jannis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347 González, José . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Grafenhorst, Matthias . . . . . . . . . . . . . . . . . . . . . 133
C Gravenkamp, Hauke . . . . . . . . . . . . . . . . . . . . . . . . 347
Grill, Maximilian J.. . . . . . . . . . . . . . . . . . . . . . . .275
Castelli, Fabian . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441 Groß, Michael . . . . . . . . . . . . . . . . . . . . 19, 129, 139
Chandrasekaran, Rajesh . . . . . . . . . . . . . . . . . . . 409 Gross, Johann. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .311
Cho, Sang Soon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Gruttmann, Friedrich. . . . . . . . . . . . . . . . . .429, 437
Concas, Francesca . . . . . . . . . . . . . . . . . . . . . . . . . . 199 Guhr, Fabian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Guo, Theron . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
D Guschke, Isabelle . . . . . . . . . . . . . . . . . . . . . . . . . . 327
Dörfler, Willy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441 H

Düster, Alexander . . . . . . . . . . . . . . . . . . . . . . . . . . 335
Damanik, Hogenrich . . . . . . . . . . . . . . . . . . . . . . . 465 Haïat, Guillaume. . . . . . . . . . . . . . . . . . . . . . . . . . .393
Diebels, Stefan . . . . . . . . . . . . . . . . . . . 199, 203, 451 Habtemariam, Abinet K. . . . . . . . . . . . . . . . . . . . 291
Dietzsch, Julian . . . . . . . . . . . . . . . . . . . 19, 129, 139 Hackl, Klaus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Dittmann, Maik . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 Hagmeyer, Nora . . . . . . . . . . . . . . . . . . . . . . 239, 247
Dittrich, Sven . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389 Hante, Stefan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Donhauser, Michael . . . . . . . . . . . . . . . . . . . . . . . . 355 Hantschke, Peter . . . . . . . . . . . . . . . . . 109, 113, 121
Dreßler, Vivian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 Hartmann, Stefan . . . . . . . . . . . . . . . . 133, 171, 191
Duong, Minh Tuan . . . . . . . . . . . . . . . . . . . . . . . . . 389 Hartung, Felix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
Duong, Thang X. . . . . . . . . . . . . . . . . . . . . . . . . . . 393 He, Qi–Chang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
479
Heider, Yousef. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .91 L
Heim, Hans–Peter . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Helmig, Jan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331 Löhnert, Stefan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Herrmann, Christoph . . . . . . . . . . . . . . . . . . . . . . 381 Lahmer, Tom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
Hesch, Christian . . . . . . . . . . . . . . . . . . . . . . . . 71, 259 Landgraf, Ralf . . . . . . . . . . . . . . . . . . . . . . . . 167, 183
Hetzler, Hartmut . . . . . . . . . . . . . . . . . . . . . . . . 83, 87 Leistner, Chris. . . . . . . . . . . . . . . . . . . . . . . . . . . . .171
Hillgärtner, Markus . . . . . . . . . . . . . . . . . . . 385, 409 Leyendecker, Sigrid . . . . . . . . . . . . . . . . . . . . . . . . 389
Hofmann, Justin Felix . . . . . . . . . . . . . . . . . . . . . 195 Lißner, Julian Richard Michael . . . . . . . . . . . . 421
Holz, David. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .389 Liao, Zisheng . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Holzapfel, Gerhard A. . . . . . . . . . . . . . . . . . . . . . . 401 Linne, Stefan. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .425
Holzinger, Markus. . . . . . . . . . . . . . . . . . . . . . . . . .445 Linse, Thomas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Hora, Pert . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Lion, Alexander . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Horn, Tobias Daniel . . . . . . . . . . . . . . . . . . . . . . . 155 Lizarazu, Jorge . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
Horny, Dominik . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 Loske, Simon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Hossain, Mokarram . . . . . . . . . . . . . . . . . . . . . . . . 159 Lund, Jorrid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
Hosters, Norbert . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
M
I
Müller, Maximilian . . . . . . . . . . . . . . . . . . . . . . . . 429
Ihlemann, Jörn . . . . . . . . . . . . . . . . . . . 155, 167, 183 Make, Michel K.P. . . . . . . . . . . . . . . . . . . . . . . . . . 243
Immel, Katharina. . . . . . . . . . . . . . . . . . . . . . . . . .393 Makvandi, Resam . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Indriyantho, Bobby Rio . . . . . . . . . . . . . . . . . . . . . 39 Manjunatha, Kiran . . . . . . . . . . . . . . . . . . . . . . . . 397
Itskov, Mikhail . . . . . . . . . . . . . . . . . . . 219, 385, 409 Markert, Bernd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Marulli, Maria Rosaria . . . . . . . . . . . . . . . . . . . . . 247
J Matzenmiller, Anton . . . . . . . . . . . . . . . . . . . 43, 355
Mayr, Matthias . . . . . . . . . . . . . . . . . . 239, 247, 263
Jin, Yan. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .51
Mehl, Miriam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
Johlitz, Michael . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Meier, Christoph . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Josyula, Siva Pavan. . . . . . . . . . . . . . . . . . . . . . . .451
Melzer, Daniel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Juhre, Daniel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .59
Menzel, Andreas . . . . . . . . . . . . . . . . . . . . . . 327, 343
Jung, Anne . . . . . . . . . . . . . . . . . . . . . . . . . . . 199, 203
Meschke, Günther . . . . . . . . . . . . . . . . . . . . . . 97, 339
Junker, Philipp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Metsch, Philipp . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
K Milow, Barbara . . . . . . . . . . . . . . . . . . . . . . . 219, 409
Miska, Niklas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Kästner, Markus . . . . . . . . . . . . . 109, 113, 121, 163 Morozovska, Anna . . . . . . . . . . . . . . . . . . . . . . . . . 459
Köllner, Anton . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 Mosler, Jörn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
König, Marcel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
Könke, Carsten . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291 N
Kühne, David . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Nahrmann, Marvin . . . . . . . . . . . . . . . . . . . . . . . . . 43
Kalina, Karl A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
Nestler, Britta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
Kaliske, Michael . . . . . . 39, 67, 105, 251, 351, 455
Nguyen, Thanh–Tung . . . . . . . . . . . . . . . . . . . . . . . 63
Kappauf, Jonas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Nguyen, Vu–Hieu . . . . . . . . . . . . . . . . . . . . . . . . . . 393
Kern, Dominik . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Ketteler, Jonas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
O
Khodor, Jad . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Kiefer, Björn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 Oliver, Javier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Kim, Myoungsic. . . . . . . . . . . . . . . . . . . . . . . . . . . .389 Ortleb, Sigrun . . . . . . . . . . . . . . . . . . . 295, 299, 303
Klassen, Markus . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
Kleemann, Susanne . . . . . . . . . . . . . . . . . . . . . . . . 425 P
Klinge, Sandra . . . . . . . . . . . . . . . . . . . . . . . . 151, 401
Klinkel, Sven . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279 Paggi, Marco . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Kochmann, Julian . . . . . . . . . . . . . . . . . . . . . . . . . . 413 Park, K.C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Kolman, Radek . . . . . . . . . . . . . . . . . . . . . . . . 23, 255 Paschkowski, Manuela . . . . . . . . . . . . . . . . . . . . . 215
Kopačka, Ján . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255 Pletz, Martin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
Krüger, Melanie . . . . . . . . . . . . . . . . . . . . . . . . . . . 259 Popa, Carmen Madalina . . . . . . . . . . . . . . . . . . . . 67
Krack, Malte . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311 Popp, Alexander . . . . . . . . 239, 247, 263, 267, 287
Kuhl, Detlef . . . . . . . . . . . . . . . . . . . . . . . . . . . 143, 195 Prager, Jens . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Kuna, Meinhard . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 Prajapati, Nishant . . . . . . . . . . . . . . . . . . . . . . . . . 381
Kunc, Oliver. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .417 Pruy, Stefan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
480
R Timothy, Jithender Jaswant . . . . . . . . . . . . . . . . 339
Titscher, Thomas . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Röbiger, Chris . . . . . . . . . . . . . . . . . . . . . . . . . . 19, 139 Tkachuk, Anton . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Römer, Ulrich J. . . . . . . . . . . . . . . . . . . . . . . . . . . . 211 Tunger, Sophie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Rütten, Markus . . . . . . . . . . . . . . . . . . . . . . . . . . . 473 Turek, Stefan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
Radtke, Lars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
Reese, Stefanie . . . . . . . . . . . . . . . . . . . . . . . . 397, 413 U
Rege, Ameya . . . . . . . . . . . . . . . . . . . . . . . . . . 219, 409
Rehbein, Thomas . . . . . . . . . . . . . . . . . . . . . . . . . . 175 Uebing, Sonja . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Reichenauer, Gudrun. . . . . . . . . . . . . . . . . . . . . . .219 Ulbrich, Stefan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
Reinold, Janis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339 Unger, Jörg F. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Reis, Martin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Ricoeur, Andreas . . . . . . . . . . . . . . . . . 179, 187, 367 V
Riesselmann, Johannes . . . . . . . . . . . . . . . . . . . . . . 27
Rosendahl, Philipp L.. . . . . . . . . . . . . . . . . . . . . .363 Völlmecke, Christina . . . . . . . . . . . . . . . . . . . . . . . . 35
Roth, Stephan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 von Boyneburgk, Claudia. . . . . . . . . . . . . . . . . . .195
Vormwald, Michael . . . . . . . . . . . . . . . . . . . . . . . . . 101
S Vorotiahin, Ivan . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
Sauer, Roger A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393 W

Schöne, Alexander . . . . . . . . . . . . . . . . . . . . . . . . . 183
Schedensack, Mira . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 Wackerfuß, Jens . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
Scheel, Johannes . . . . . . . . . . . . . . . . . . . . . . . . . . . 367 Waimann, Johanna . . . . . . . . . . . . . . . . . . . . . . 7, 413
Scheunemann, Lisa . . . . . . . . . . . . . . . . . . . . . . . . . 235 Waldmann, Daniele . . . . . . . . . . . . . . . . . . . . . . . . . 63
Schleifer, Felix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445 Wall, Wolfgang A. . . . . . . . . . . . . . . . . . . . . . . . . . 275
Schlosser, Alexander . . . . . . . . . . . . . . . . . . . . . . . 179 Walloth, Mirjam . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
Schneider, Alexandra P. . . . . . . . . . . . . . . . . . . . . 363 Warkentin, Andreas . . . . . . . . . . . . . . . . . . . . . . . 187
Schneider, Daniel . . . . . . . . . . . . . . . . . . . . . . . . . . 381 Waschinsky, Navina . . . . . . . . . . . . . . . . . . . . . . . 343
Schneider, Jens. . . . . . . . . . . . . . . . . . . . . . . . . . . . .363 Wasmer, Paul . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Schröder, Bettina . . . . . . . . . . . . . . . . . . . . . . . . . . 143 Wick, Thomas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Schröder, Jörg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235 Wiegard, Bjarne . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
Schuecker, Clara . . . . . . . . . . . . . . . . . . . . . . . . . . . 371 Wiegold, Tillmann . . . . . . . . . . . . . . . . . . . . . . . . . 401
Schulte, Jonathan . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 Williams Moises, Rodolfo Javier. . . . . . . . . . . . .97
Schulz, Katrin . . . . . . . . . . . . . . . . . . . . . . . . . . 47, 227 Wohlgemuth, Felix. . . . . . . . . . . . . . . . . . . . . . . . .375
Seidl, Marina . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377 Wohlmuth, Barbara . . . . . . . . . . . . . . . . . . . . . . . . 267
Seiler, Martha . . . . . . . . . . . . . . . . . . . . . . . . 109, 121 Wollny, Ines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
Seitz, Alexander . . . . . . . . . . . . . . . . . . . . . . . . . . . 267 Wriggers, Peter. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .75
Selzer, Michael . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381 Wulf, Hans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Sewerin, Fabian. . . . . . . . . . . . . . . . . . . . . . . . . . . .231 Wunderlich, Linus . . . . . . . . . . . . . . . . . . . . . . . . . 267
Shekarchizadeh, Navid . . . . . . . . . . . . . . . . . . . . . . 31
Souli, Mhamed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377 X
Späth, Michael . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
Spanos, Pol D. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 Xu, Bai–Xiang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
Staudt, Yves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
Steinbrecher, Ivo. . . . . . . . . . . . . . . . . . . . . . . . . . .287 Y
Storm, Johannes . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
Straub, Veronika. . . . . . . . . . . . . . . . . . . . . . . . . . .303 Yvonnet, Julien . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Sun, Wai Ching . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
Supriatna, Dennie . . . . . . . . . . . . . . . . . . . . . . . . . 455 Z
Svendsen, Bob . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
Zabel, Volkmar. . . . . . . . . . . . . . . . . . . . . . . . . . . . .291
Sweidan, Abdel Hassan . . . . . . . . . . . . . . . . . . . . . 91
Zahn, Anna . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
T Zeifang, Jonas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Zheng, Qige . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
Töller, Felix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 Zhou, Xian–Dong . . . . . . . . . . . . . . . . . . . . . . . . . . 463
Tartaglione, Fabiola . . . . . . . . . . . . . . . . . . . . . . . 291 Zoller, Jan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
Tauscher, Markus. . . . . . . . . . . . . . . . . . . . . . . . . .371 Zoller, Kolja . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

Online - OpenAccess - 978 3 7376 5093 9.OpenAccess

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Online - OpenAccess - 978 3 7376 5093 9.OpenAccess

Uploaded by

Copyright:

Available Formats

8th GACM Colloquium on

For Young Scientists From Academia and Industry

August 28th – 30th , 2019

University of Kassel, Germany

Bibliographic information published by Deutsche Nationalbibliothek

© 2019, kassel university press GmbH, Kassel

Kassel, August 2019

GACM Best PhD Award Winner Lectures 1

MS 01: Applications and Computational Methods in Metamaterials 11

MS 02: Peridynamics and other non–Local Methods for Computational Mechanics 39

MS 03: Current Trends in Computational Fracture Mechanics 51

MS 04: Modern Trends in Numerical Continuation of Stationary Solutions and Application 79

MS 05: Multi-field problems in porous media mechanics 91

MS 07: Higher Order Numerical Methods in Space and Time 129

MS 08: Smart, Active and Polymer materials 151

MS 12: Computational Multiscale and Multiphysics Simulations of Microheterogeneous Materials219

MS 13: Advanced Numerical Methods for Fluid-Structure Interaction 239

MS 14: Computational Contact & Interface Mechanics 247

MS 17: Coupled Problems 311

MS 19: Brittle Fracture Analysis in Solids: Experiment and Simulation 377

MS 20: Numerical methods in biomechanics 385

MS 21: Representative Volume Elements: Microstructure Analysis and Homogenization 409

MS 23: Non–Newtonian Fluids 465

Authors Index 479

Proceedings of the 8th GACM Colloquium on Computational Mechanics

The Mechanics of Brain Development

Computational model for brain growth

∗ Corresponding author: silvia.budday@fau.de ORCID iD: 0000-0002-7072-8174

Experimental model for brain growth

Proceedings of the 8th GACM Colloquium on Computational Mechanics

Kino-Geometric Modeling of Proteins and Molecular

KGS – our Kino-Geometric Sampling framework

∗ Corresponding author: dominik.budday@fau.de ORCID iD: ORCID 0000-0002-6180-2125

Results and Discussion

Proceedings of the 8th GACM Colloquium on Computational Mechanics

Variational modeling of irreversible eects in cycled

1 Institute of Applied Mechanics, RWTH Aachen University, Germany

1 Simultaneous formation of irreversible martensite

∗ Corresponding author: waimann@ifam.rwth-aachen.de ORCID iD: 0000-0002-7579-7240

2 Simultaneous plastic evolution

3 Storage of the transformation history

Proceedings of the 8th GACM Colloquium on Computational Mechanics

Connection between micro- and macroscales

Strain gradient theory and a computational comparison verifying its necessity

∗ Corresponding author: bilenemek@abali.org ORCID iD: 0000-0002-8735-6071

Figure 1. Microscale description of the structure, B, with two mate-

Proceedings of the 8th GACM Colloquium on Computational Mechanics

On the modeling of microstructure influence

It is well known that linear elasticity obeys the Hooke law

utt = c2 uxx . (7)

The equation of motion for the strain gradient model is extended to

ρutt = ρc2 uxx + Aϕx , (9)

2.1 Dimensionless variables

For further analysis we introduce dimensionless variables

we have for the microstructure model with internal variables

UT T = UXX + aΦX , (13)

Again, the choice b = −2a leads to the Mindlin microelasticity.

−Û ω 2 = −Û k 2 + aΦ̂(ik), (16)

This can be represented in the matrix form

Figure 1. Dispersion curves corresponding to Eq. (21).

UT T = UXX + aΦX , (22)

Figure 2. Dispersion curves corresponding to Eq. (24).

Proceedings of the 8th GACM Colloquium on Computational Mechanics

Variational modeling of irreversible eects in cycled

A nite strain microplane-plasticity model for concrete